mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-08-26 18:17:52 +02:00
Compare commits
47
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
 Georgi Gerganov
|
e7fbfc9b80 | ||
|
Andreas KieslingerandGeorgi Gerganov
|
d46bd7ef2d | ||
|
Andreas KieslingerandGeorgi Gerganov
|
070933684f | ||
|
aendkandGeorgi Gerganov
|
1528c841dc | ||
|
aendkandGeorgi Gerganov
|
ff28ae93a2 | ||
|
aendkandGeorgi Gerganov
|
01d89f9b96 | ||
|
aendkandGeorgi Gerganov
|
e74b070e30 | ||
|
aendkandGeorgi Gerganov
|
b7376c3ed7 | ||
|
aendkandGeorgi Gerganov
|
d776354dc9 | ||
|
aendkandGeorgi Gerganov
|
79a77277ad | ||
|
aendkandGeorgi Gerganov
|
44e481bb34 | ||
|
aendkandGeorgi Gerganov
|
91c6026b5c | ||
|
aendkandGeorgi Gerganov
|
2ad0d391e1 | ||
|
aendkandGeorgi Gerganov
|
dd9f1faf42 | ||
|
Georgi Gerganov
|
a554bdd70f | ||
|
Johannes Gäßler
|
02ee504f90 | ||
|
Johannes Gäßler
|
94c66557b8 | ||
|
Johannes Gäßler
|
6ab02d0908 | ||
|
Johannes Gäßler
|
2b7055381c | ||
|
Johannes Gäßler
|
b7fd10664e | ||
|
Johannes Gäßler
|
a630b27da7 | ||
|
Johannes Gäßler
|
39b96f8fe1 | ||
|
Georgi GerganovandGitHub
|
22cae83218 | ||
|
Jeff BolzandGitHub
|
449ec2ab07 | ||
|
Georgi GerganovandGitHub
|
3795cc1e89 | ||
|
Sigbjørn SkjæretandGitHub
|
b828e18c75 | ||
|
Adrien GallouëtandGitHub
|
a4ea7a188f | ||
|
Georgi GerganovandGitHub
|
7a4f97d196 | ||
|
Oleksandr KuvshynovandGitHub
|
a498c75ad1 | ||
|
Jeff BolzandGitHub
|
3409ab842d | ||
|
Jeff BolzandGitHub
|
c342c3b93d | ||
|
will-lmsandGitHub
|
af252d0758 | ||
|
Sigbjørn SkjæretandGitHub
|
11fb327bf3 | ||
|
Aaron TeoandGitHub
|
e6e934c5ea | ||
|
|
b536eb0233 | ||
|
Xuan-Son NguyenandGitHub
|
e0c93af2a0 | ||
|
Georgi GerganovandGitHub
|
423bee462b | ||
|
Xuan-Son NguyenandGitHub
|
8abcc70a74 | ||
|
|
eaba92c3dc | ||
|
Daniel BeveniusandGitHub
|
6ab881b7c3 | ||
|
Georgi GerganovandGitHub
|
d838c22bb3 | ||
|
Daniel BeveniusandGitHub
|
25f40ca65f | ||
|
Kevin PougetandGitHub
|
015deb9048 | ||
|
Aman GuptaandGitHub
|
2ceda3f662 | ||
|
Georgi GerganovandGitHub
|
44008ce8f9 | ||
|
Georgi GerganovandGitHub
|
6a9bf2f788 | ||
|
Georgi GerganovandGitHub
|
faa1bc26ee |
@@ -54,6 +54,7 @@ RUN apt-get update \
|
||||
build-essential \
|
||||
git \
|
||||
python3 \
|
||||
python3-dev \
|
||||
python3-pip \
|
||||
python3-wheel \
|
||||
&& pip install --break-system-packages --upgrade setuptools \
|
||||
|
||||
@@ -293,6 +293,7 @@ jobs:
|
||||
cmake -B build \
|
||||
-DLLAMA_FATAL_WARNINGS=ON \
|
||||
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
|
||||
|
||||
@@ -303,6 +304,7 @@ jobs:
|
||||
cmake -B build \
|
||||
-DLLAMA_FATAL_WARNINGS=ON \
|
||||
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
|
||||
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
|
||||
-DGGML_OPENMP=OFF
|
||||
cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
|
||||
|
||||
@@ -36,7 +36,7 @@ jobs:
|
||||
|
||||
strategy:
|
||||
matrix:
|
||||
sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
|
||||
sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
|
||||
build_type: [RelWithDebInfo]
|
||||
include:
|
||||
- build_type: Release
|
||||
@@ -45,7 +45,7 @@ jobs:
|
||||
- build_type: Release
|
||||
sanitizer: ""
|
||||
extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
|
||||
fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
|
||||
fail-fast: false
|
||||
|
||||
steps:
|
||||
- name: Dependencies
|
||||
@@ -72,7 +72,15 @@ jobs:
|
||||
- name: Build
|
||||
id: cmake_build
|
||||
run: |
|
||||
cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
|
||||
cmake -B build \
|
||||
-DLLAMA_BUILD_BORINGSSL=ON \
|
||||
-DGGML_SCHED_NO_REALLOC=ON \
|
||||
-DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
|
||||
-DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
|
||||
-DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
|
||||
-DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
|
||||
-DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
|
||||
-DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
|
||||
cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
|
||||
|
||||
- name: Python setup
|
||||
@@ -88,7 +96,7 @@ jobs:
|
||||
|
||||
- name: Tests
|
||||
id: server_integration_tests
|
||||
if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) && matrix.build_type == 'Release' }}
|
||||
if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
|
||||
run: |
|
||||
cd tools/server/tests
|
||||
export ${{ matrix.extra_args }}
|
||||
|
||||
@@ -164,29 +164,6 @@ llama_option_depr(WARNING LLAMA_SYCL GGML_SYCL)
|
||||
llama_option_depr(WARNING LLAMA_SYCL_F16 GGML_SYCL_F16)
|
||||
llama_option_depr(WARNING LLAMA_CANN GGML_CANN)
|
||||
|
||||
if (NOT MSVC)
|
||||
if (LLAMA_SANITIZE_THREAD)
|
||||
message(STATUS "Using -fsanitize=thread")
|
||||
|
||||
add_compile_options(-fsanitize=thread)
|
||||
link_libraries (-fsanitize=thread)
|
||||
endif()
|
||||
|
||||
if (LLAMA_SANITIZE_ADDRESS)
|
||||
message(STATUS "Using -fsanitize=address")
|
||||
|
||||
add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
|
||||
link_libraries (-fsanitize=address)
|
||||
endif()
|
||||
|
||||
if (LLAMA_SANITIZE_UNDEFINED)
|
||||
message(STATUS "Using -fsanitize=undefined")
|
||||
|
||||
add_compile_options(-fsanitize=undefined)
|
||||
link_libraries (-fsanitize=undefined)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
include("cmake/license.cmake")
|
||||
license_add_file("llama.cpp" "LICENSE")
|
||||
|
||||
|
||||
@@ -27,6 +27,7 @@
|
||||
/examples/batched.swift/ @ggerganov
|
||||
/examples/batched/ @ggerganov
|
||||
/examples/convert-llama2c-to-ggml/ @ggerganov
|
||||
/examples/debug/ @danbev @pwilkin
|
||||
/examples/deprecation-warning/ @ggerganov
|
||||
/examples/diffusion/ @am17an
|
||||
/examples/embedding/ @ggerganov
|
||||
|
||||
+209
-162
@@ -8,7 +8,7 @@ g++ --version
|
||||
g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
|
||||
|
||||
nvidia-smi
|
||||
Sun Nov 2 10:43:25 2025
|
||||
Thu Feb 5 13:49:40 2026
|
||||
+-----------------------------------------------------------------------------------------+
|
||||
| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 |
|
||||
+-----------------------------------------+------------------------+----------------------+
|
||||
@@ -17,7 +17,7 @@ Sun Nov 2 10:43:25 2025
|
||||
| | | MIG M. |
|
||||
|=========================================+========================+======================|
|
||||
| 0 NVIDIA GB10 On | 0000000F:01:00.0 Off | N/A |
|
||||
| N/A 35C P8 4W / N/A | Not Supported | 0% Default |
|
||||
| N/A 47C P0 13W / N/A | Not Supported | 0% Default |
|
||||
| | | N/A |
|
||||
+-----------------------------------------+------------------------+----------------------+
|
||||
```
|
||||
@@ -29,46 +29,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.374 | 1369.01 | 0.383 | 83.64 | 0.757 | 719.01 |
|
||||
| 512 | 32 | 2 | 1088 | 0.274 | 3741.35 | 0.659 | 97.14 | 0.933 | 1166.66 |
|
||||
| 512 | 32 | 4 | 2176 | 0.526 | 3896.47 | 0.817 | 156.73 | 1.342 | 1621.08 |
|
||||
| 512 | 32 | 8 | 4352 | 1.044 | 3925.10 | 0.987 | 259.44 | 2.030 | 2143.56 |
|
||||
| 512 | 32 | 16 | 8704 | 2.076 | 3945.84 | 1.248 | 410.32 | 3.324 | 2618.60 |
|
||||
| 512 | 32 | 32 | 17408 | 4.170 | 3929.28 | 1.630 | 628.40 | 5.799 | 3001.76 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.083 | 3782.66 | 0.394 | 81.21 | 1.477 | 2795.13 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.166 | 3782.72 | 0.725 | 88.28 | 2.891 | 2856.14 |
|
||||
| 4096 | 32 | 4 | 16512 | 4.333 | 3780.88 | 0.896 | 142.82 | 5.230 | 3157.38 |
|
||||
| 4096 | 32 | 8 | 33024 | 8.618 | 3802.14 | 1.155 | 221.69 | 9.773 | 3379.08 |
|
||||
| 4096 | 32 | 16 | 66048 | 17.330 | 3781.73 | 1.598 | 320.34 | 18.928 | 3489.45 |
|
||||
| 4096 | 32 | 32 | 132096 | 34.671 | 3780.48 | 2.336 | 438.35 | 37.007 | 3569.51 |
|
||||
| 8192 | 32 | 1 | 8224 | 2.233 | 3668.56 | 0.438 | 72.98 | 2.671 | 3078.44 |
|
||||
| 8192 | 32 | 2 | 16448 | 4.425 | 3702.95 | 0.756 | 84.66 | 5.181 | 3174.95 |
|
||||
| 8192 | 32 | 4 | 32896 | 8.859 | 3698.64 | 0.967 | 132.38 | 9.826 | 3347.72 |
|
||||
| 8192 | 32 | 8 | 65792 | 17.714 | 3699.57 | 1.277 | 200.52 | 18.991 | 3464.35 |
|
||||
| 8192 | 32 | 16 | 131584 | 35.494 | 3692.84 | 1.841 | 278.12 | 37.335 | 3524.46 |
|
||||
| 8192 | 32 | 32 | 263168 | 70.949 | 3694.82 | 2.798 | 365.99 | 73.747 | 3568.53 |
|
||||
| 512 | 32 | 1 | 544 | 0.270 | 1895.57 | 0.399 | 80.13 | 0.669 | 812.60 |
|
||||
| 512 | 32 | 2 | 1088 | 0.230 | 4451.23 | 0.583 | 109.71 | 0.813 | 1337.56 |
|
||||
| 512 | 32 | 4 | 2176 | 0.437 | 4688.87 | 0.820 | 156.03 | 1.257 | 1730.91 |
|
||||
| 512 | 32 | 8 | 4352 | 0.863 | 4744.23 | 0.942 | 271.79 | 1.805 | 2410.73 |
|
||||
| 512 | 32 | 16 | 8704 | 1.725 | 4748.19 | 1.173 | 436.38 | 2.899 | 3002.85 |
|
||||
| 512 | 32 | 32 | 17408 | 3.437 | 4767.38 | 1.503 | 681.49 | 4.939 | 3524.40 |
|
||||
| 4096 | 32 | 1 | 4128 | 0.907 | 4513.91 | 0.407 | 78.54 | 1.315 | 3139.56 |
|
||||
| 4096 | 32 | 2 | 8256 | 1.796 | 4560.42 | 0.625 | 102.37 | 2.422 | 3409.45 |
|
||||
| 4096 | 32 | 4 | 16512 | 3.596 | 4555.66 | 0.888 | 144.11 | 4.485 | 3681.93 |
|
||||
| 4096 | 32 | 8 | 33024 | 7.184 | 4561.44 | 1.098 | 233.11 | 8.282 | 3987.51 |
|
||||
| 4096 | 32 | 16 | 66048 | 14.369 | 4560.82 | 1.503 | 340.74 | 15.872 | 4161.30 |
|
||||
| 4096 | 32 | 32 | 132096 | 28.760 | 4557.52 | 2.162 | 473.59 | 30.922 | 4271.95 |
|
||||
| 8192 | 32 | 1 | 8224 | 1.859 | 4405.59 | 0.430 | 74.36 | 2.290 | 3591.61 |
|
||||
| 8192 | 32 | 2 | 16448 | 3.698 | 4430.02 | 0.656 | 97.59 | 4.354 | 3777.47 |
|
||||
| 8192 | 32 | 4 | 32896 | 7.403 | 4426.10 | 0.957 | 133.82 | 8.360 | 3934.97 |
|
||||
| 8192 | 32 | 8 | 65792 | 14.802 | 4427.63 | 1.222 | 209.44 | 16.024 | 4105.87 |
|
||||
| 8192 | 32 | 16 | 131584 | 29.596 | 4428.67 | 1.741 | 294.13 | 31.337 | 4199.00 |
|
||||
| 8192 | 32 | 32 | 263168 | 59.169 | 4430.42 | 2.619 | 390.92 | 61.789 | 4259.17 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 3714.25 ± 20.36 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 86.58 ± 0.43 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 3445.17 ± 17.85 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 81.72 ± 0.53 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 3218.78 ± 11.34 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 74.86 ± 0.64 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 2732.83 ± 7.17 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 71.57 ± 0.51 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 2119.75 ± 12.81 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 62.33 ± 0.24 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 4505.82 ± 12.90 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 83.43 ± 0.59 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 4158.34 ± 18.84 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 79.22 ± 0.60 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 3993.81 ± 17.55 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 75.22 ± 1.05 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 3449.98 ± 12.13 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 70.36 ± 0.37 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 2689.42 ± 18.89 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 61.65 ± 0.30 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/gpt-oss-120b-GGUF
|
||||
|
||||
@@ -77,46 +77,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.571 | 897.18 | 0.543 | 58.96 | 1.113 | 488.60 |
|
||||
| 512 | 32 | 2 | 1088 | 0.593 | 1725.37 | 1.041 | 61.45 | 1.635 | 665.48 |
|
||||
| 512 | 32 | 4 | 2176 | 1.043 | 1963.15 | 1.334 | 95.95 | 2.377 | 915.36 |
|
||||
| 512 | 32 | 8 | 4352 | 2.099 | 1951.63 | 1.717 | 149.07 | 3.816 | 1140.45 |
|
||||
| 512 | 32 | 16 | 8704 | 4.207 | 1947.12 | 2.311 | 221.56 | 6.518 | 1335.35 |
|
||||
| 512 | 32 | 32 | 17408 | 8.422 | 1945.36 | 3.298 | 310.46 | 11.720 | 1485.27 |
|
||||
| 4096 | 32 | 1 | 4128 | 2.138 | 1915.88 | 0.571 | 56.09 | 2.708 | 1524.12 |
|
||||
| 4096 | 32 | 2 | 8256 | 4.266 | 1920.25 | 1.137 | 56.27 | 5.404 | 1527.90 |
|
||||
| 4096 | 32 | 4 | 16512 | 8.564 | 1913.02 | 1.471 | 86.99 | 10.036 | 1645.29 |
|
||||
| 4096 | 32 | 8 | 33024 | 17.092 | 1917.19 | 1.979 | 129.33 | 19.071 | 1731.63 |
|
||||
| 4096 | 32 | 16 | 66048 | 34.211 | 1915.65 | 2.850 | 179.66 | 37.061 | 1782.15 |
|
||||
| 4096 | 32 | 32 | 132096 | 68.394 | 1916.44 | 4.381 | 233.72 | 72.775 | 1815.13 |
|
||||
| 8192 | 32 | 1 | 8224 | 4.349 | 1883.45 | 0.620 | 51.65 | 4.969 | 1655.04 |
|
||||
| 8192 | 32 | 2 | 16448 | 8.674 | 1888.83 | 1.178 | 54.33 | 9.852 | 1669.48 |
|
||||
| 8192 | 32 | 4 | 32896 | 17.351 | 1888.55 | 1.580 | 81.01 | 18.931 | 1737.68 |
|
||||
| 8192 | 32 | 8 | 65792 | 34.743 | 1886.31 | 2.173 | 117.80 | 36.916 | 1782.20 |
|
||||
| 8192 | 32 | 16 | 131584 | 69.413 | 1888.29 | 3.297 | 155.28 | 72.710 | 1809.70 |
|
||||
| 8192 | 32 | 32 | 263168 | 138.903 | 1887.24 | 5.004 | 204.63 | 143.907 | 1828.73 |
|
||||
| 512 | 32 | 1 | 544 | 0.445 | 1151.80 | 0.560 | 57.14 | 1.005 | 541.53 |
|
||||
| 512 | 32 | 2 | 1088 | 0.472 | 2169.85 | 0.874 | 73.27 | 1.345 | 808.65 |
|
||||
| 512 | 32 | 4 | 2176 | 0.826 | 2480.33 | 1.299 | 98.51 | 2.125 | 1023.94 |
|
||||
| 512 | 32 | 8 | 4352 | 1.644 | 2491.67 | 1.608 | 159.18 | 3.252 | 1338.20 |
|
||||
| 512 | 32 | 16 | 8704 | 3.292 | 2488.35 | 2.117 | 241.85 | 5.409 | 1609.13 |
|
||||
| 512 | 32 | 32 | 17408 | 6.604 | 2481.07 | 2.898 | 353.31 | 9.502 | 1832.04 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.698 | 2412.65 | 0.580 | 55.21 | 2.277 | 1812.66 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.399 | 2409.88 | 0.934 | 68.53 | 4.333 | 1905.27 |
|
||||
| 4096 | 32 | 4 | 16512 | 6.823 | 2401.21 | 1.411 | 90.72 | 8.234 | 2005.30 |
|
||||
| 4096 | 32 | 8 | 33024 | 13.574 | 2413.97 | 1.841 | 139.07 | 15.415 | 2142.31 |
|
||||
| 4096 | 32 | 16 | 66048 | 27.176 | 2411.52 | 2.609 | 196.26 | 29.785 | 2217.49 |
|
||||
| 4096 | 32 | 32 | 132096 | 54.359 | 2411.23 | 3.905 | 262.20 | 58.264 | 2267.19 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.491 | 2346.81 | 0.613 | 52.23 | 4.103 | 2004.21 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.939 | 2361.03 | 0.981 | 65.21 | 7.921 | 2076.56 |
|
||||
| 8192 | 32 | 4 | 32896 | 13.888 | 2359.40 | 1.511 | 84.71 | 15.399 | 2136.21 |
|
||||
| 8192 | 32 | 8 | 65792 | 27.756 | 2361.18 | 2.034 | 125.86 | 29.790 | 2208.56 |
|
||||
| 8192 | 32 | 16 | 131584 | 55.554 | 2359.34 | 3.021 | 169.49 | 58.575 | 2246.41 |
|
||||
| 8192 | 32 | 32 | 263168 | 111.036 | 2360.89 | 4.537 | 225.72 | 115.573 | 2277.08 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 1919.36 ± 5.01 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 60.40 ± 0.30 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 1825.30 ± 6.37 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 56.94 ± 0.29 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 1739.19 ± 6.00 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 52.51 ± 0.42 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 1536.75 ± 4.27 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 49.33 ± 0.27 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 1255.85 ± 3.26 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 42.99 ± 0.18 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 2443.91 ± 7.47 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 58.72 ± 0.20 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 2309.84 ± 3.63 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 55.67 ± 0.35 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 2216.68 ± 10.16 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 52.87 ± 0.43 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 1956.31 ± 6.39 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 49.45 ± 0.20 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 1567.08 ± 11.79 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 42.76 ± 0.14 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
|
||||
@@ -125,46 +125,46 @@ Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.398 | 1285.90 | 0.530 | 60.41 | 0.928 | 586.27 |
|
||||
| 512 | 32 | 2 | 1088 | 0.386 | 2651.65 | 0.948 | 67.50 | 1.334 | 815.38 |
|
||||
| 512 | 32 | 4 | 2176 | 0.666 | 3076.37 | 1.209 | 105.87 | 1.875 | 1160.71 |
|
||||
| 512 | 32 | 8 | 4352 | 1.325 | 3091.39 | 1.610 | 158.98 | 2.935 | 1482.65 |
|
||||
| 512 | 32 | 16 | 8704 | 2.664 | 3075.58 | 2.150 | 238.19 | 4.813 | 1808.39 |
|
||||
| 512 | 32 | 32 | 17408 | 5.336 | 3070.31 | 2.904 | 352.59 | 8.240 | 2112.50 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.444 | 2836.81 | 0.581 | 55.09 | 2.025 | 2038.81 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.872 | 2852.14 | 1.084 | 59.06 | 3.956 | 2086.99 |
|
||||
| 4096 | 32 | 4 | 16512 | 5.744 | 2852.32 | 1.440 | 88.90 | 7.184 | 2298.47 |
|
||||
| 4096 | 32 | 8 | 33024 | 11.463 | 2858.68 | 2.068 | 123.78 | 13.531 | 2440.65 |
|
||||
| 4096 | 32 | 16 | 66048 | 22.915 | 2859.95 | 3.018 | 169.67 | 25.933 | 2546.90 |
|
||||
| 4096 | 32 | 32 | 132096 | 45.956 | 2852.10 | 4.609 | 222.18 | 50.565 | 2612.39 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.063 | 2674.72 | 0.693 | 46.20 | 3.755 | 2189.92 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.109 | 2681.87 | 1.214 | 52.71 | 7.323 | 2245.98 |
|
||||
| 8192 | 32 | 4 | 32896 | 12.197 | 2686.63 | 1.682 | 76.11 | 13.878 | 2370.30 |
|
||||
| 8192 | 32 | 8 | 65792 | 24.409 | 2684.94 | 2.556 | 100.17 | 26.965 | 2439.95 |
|
||||
| 8192 | 32 | 16 | 131584 | 48.753 | 2688.50 | 3.994 | 128.20 | 52.747 | 2494.64 |
|
||||
| 8192 | 32 | 32 | 263168 | 97.508 | 2688.42 | 6.528 | 156.86 | 104.037 | 2529.57 |
|
||||
| 512 | 32 | 1 | 544 | 0.393 | 1303.73 | 0.548 | 58.36 | 0.941 | 578.10 |
|
||||
| 512 | 32 | 2 | 1088 | 0.387 | 2648.68 | 0.910 | 70.35 | 1.296 | 839.27 |
|
||||
| 512 | 32 | 4 | 2176 | 0.659 | 3107.63 | 1.302 | 98.33 | 1.961 | 1109.77 |
|
||||
| 512 | 32 | 8 | 4352 | 1.322 | 3099.35 | 1.669 | 153.42 | 2.990 | 1455.43 |
|
||||
| 512 | 32 | 16 | 8704 | 2.639 | 3104.63 | 2.212 | 231.44 | 4.851 | 1794.32 |
|
||||
| 512 | 32 | 32 | 17408 | 5.284 | 3100.80 | 2.955 | 346.53 | 8.239 | 2112.93 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.417 | 2890.36 | 0.598 | 53.51 | 2.015 | 2048.45 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.829 | 2895.62 | 1.019 | 62.82 | 3.848 | 2145.60 |
|
||||
| 4096 | 32 | 4 | 16512 | 5.656 | 2896.96 | 1.528 | 83.79 | 7.183 | 2298.71 |
|
||||
| 4096 | 32 | 8 | 33024 | 11.338 | 2890.02 | 2.127 | 120.36 | 13.465 | 2452.53 |
|
||||
| 4096 | 32 | 16 | 66048 | 22.709 | 2885.96 | 3.104 | 164.97 | 25.812 | 2558.79 |
|
||||
| 4096 | 32 | 32 | 132096 | 45.301 | 2893.35 | 4.723 | 216.80 | 50.024 | 2640.63 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.022 | 2711.09 | 0.678 | 47.20 | 3.700 | 2222.89 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.039 | 2713.01 | 1.149 | 55.70 | 7.188 | 2288.21 |
|
||||
| 8192 | 32 | 4 | 32896 | 12.050 | 2719.35 | 1.785 | 71.69 | 13.835 | 2377.67 |
|
||||
| 8192 | 32 | 8 | 65792 | 24.113 | 2717.90 | 2.629 | 97.39 | 26.741 | 2460.31 |
|
||||
| 8192 | 32 | 16 | 131584 | 48.178 | 2720.58 | 4.099 | 124.91 | 52.277 | 2517.06 |
|
||||
| 8192 | 32 | 32 | 263168 | 96.401 | 2719.31 | 6.696 | 152.93 | 103.097 | 2552.63 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 2925.55 ± 4.25 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 62.80 ± 0.27 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 2531.01 ± 6.79 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 55.86 ± 0.33 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 2244.39 ± 5.33 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 45.95 ± 0.33 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 1783.17 ± 3.68 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 39.07 ± 0.10 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 1241.90 ± 3.13 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 29.92 ± 0.06 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 2986.97 ± 18.87 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 61.06 ± 0.23 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 2633.45 ± 6.26 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 54.77 ± 0.28 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 2354.14 ± 3.84 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 48.02 ± 0.40 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 1908.86 ± 4.25 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 40.23 ± 0.10 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 1348.17 ± 2.00 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 30.21 ± 0.04 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
|
||||
@@ -173,46 +173,46 @@ Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.211 | 2421.57 | 1.055 | 30.33 | 1.266 | 429.57 |
|
||||
| 512 | 32 | 2 | 1088 | 0.419 | 2441.34 | 1.130 | 56.65 | 1.549 | 702.32 |
|
||||
| 512 | 32 | 4 | 2176 | 0.873 | 2345.54 | 1.174 | 108.99 | 2.048 | 1062.74 |
|
||||
| 512 | 32 | 8 | 4352 | 1.727 | 2371.85 | 1.254 | 204.22 | 2.980 | 1460.19 |
|
||||
| 512 | 32 | 16 | 8704 | 3.452 | 2373.22 | 1.492 | 343.16 | 4.944 | 1760.56 |
|
||||
| 512 | 32 | 32 | 17408 | 6.916 | 2368.93 | 1.675 | 611.51 | 8.591 | 2026.36 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.799 | 2277.26 | 1.084 | 29.51 | 2.883 | 1431.91 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.577 | 2290.01 | 1.196 | 53.50 | 4.774 | 1729.51 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.172 | 2284.36 | 1.313 | 97.50 | 8.485 | 1946.00 |
|
||||
| 4096 | 32 | 8 | 33024 | 14.341 | 2284.96 | 1.520 | 168.46 | 15.860 | 2082.18 |
|
||||
| 4096 | 32 | 16 | 66048 | 28.675 | 2285.44 | 1.983 | 258.21 | 30.658 | 2154.33 |
|
||||
| 4096 | 32 | 32 | 132096 | 57.354 | 2285.32 | 2.640 | 387.87 | 59.994 | 2201.82 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.701 | 2213.75 | 1.119 | 28.59 | 4.820 | 1706.34 |
|
||||
| 8192 | 32 | 2 | 16448 | 7.410 | 2211.19 | 1.272 | 50.31 | 8.682 | 1894.56 |
|
||||
| 8192 | 32 | 4 | 32896 | 14.802 | 2213.83 | 1.460 | 87.68 | 16.261 | 2022.96 |
|
||||
| 8192 | 32 | 8 | 65792 | 29.609 | 2213.35 | 1.781 | 143.74 | 31.390 | 2095.93 |
|
||||
| 8192 | 32 | 16 | 131584 | 59.229 | 2212.96 | 2.495 | 205.17 | 61.725 | 2131.79 |
|
||||
| 8192 | 32 | 32 | 263168 | 118.449 | 2213.15 | 3.714 | 275.75 | 122.162 | 2154.25 |
|
||||
| 512 | 32 | 1 | 544 | 0.212 | 2420.12 | 1.100 | 29.10 | 1.311 | 414.85 |
|
||||
| 512 | 32 | 2 | 1088 | 0.428 | 2393.89 | 1.185 | 54.00 | 1.613 | 674.56 |
|
||||
| 512 | 32 | 4 | 2176 | 0.894 | 2290.41 | 1.229 | 104.17 | 2.123 | 1025.02 |
|
||||
| 512 | 32 | 8 | 4352 | 1.758 | 2330.36 | 1.319 | 194.15 | 3.076 | 1414.70 |
|
||||
| 512 | 32 | 16 | 8704 | 3.508 | 2335.21 | 1.543 | 331.90 | 5.051 | 1723.33 |
|
||||
| 512 | 32 | 32 | 17408 | 7.035 | 2328.93 | 1.738 | 589.21 | 8.773 | 1984.29 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.831 | 2237.25 | 1.125 | 28.44 | 2.956 | 1396.42 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.642 | 2249.48 | 1.253 | 51.07 | 4.895 | 1686.64 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.274 | 2252.26 | 1.380 | 92.72 | 8.655 | 1907.81 |
|
||||
| 4096 | 32 | 8 | 33024 | 14.576 | 2248.09 | 1.617 | 158.29 | 16.193 | 2039.37 |
|
||||
| 4096 | 32 | 16 | 66048 | 29.138 | 2249.17 | 2.081 | 246.01 | 31.219 | 2115.63 |
|
||||
| 4096 | 32 | 32 | 132096 | 58.275 | 2249.19 | 2.814 | 363.87 | 61.089 | 2162.34 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.757 | 2180.26 | 1.184 | 27.03 | 4.941 | 1664.37 |
|
||||
| 8192 | 32 | 2 | 16448 | 7.522 | 2178.05 | 1.341 | 47.73 | 8.863 | 1855.77 |
|
||||
| 8192 | 32 | 4 | 32896 | 15.043 | 2178.25 | 1.548 | 82.69 | 16.591 | 1982.74 |
|
||||
| 8192 | 32 | 8 | 65792 | 30.111 | 2176.49 | 1.937 | 132.13 | 32.048 | 2052.90 |
|
||||
| 8192 | 32 | 16 | 131584 | 60.405 | 2169.90 | 2.706 | 189.21 | 63.111 | 2084.97 |
|
||||
| 8192 | 32 | 32 | 263168 | 120.439 | 2176.58 | 3.993 | 256.46 | 124.432 | 2114.96 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 2272.74 ± 4.68 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 30.66 ± 0.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 2107.80 ± 9.55 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 29.71 ± 0.05 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 1937.80 ± 6.75 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 28.86 ± 0.04 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 1641.12 ± 1.78 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 27.24 ± 0.04 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 1296.02 ± 2.67 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 23.78 ± 0.03 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 2250.28 ± 6.41 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 29.43 ± 0.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 2100.19 ± 8.96 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 28.61 ± 0.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 2007.56 ± 4.16 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 27.38 ± 0.09 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 1779.11 ± 6.42 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 25.72 ± 0.03 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 1471.23 ± 1.71 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 22.51 ± 0.02 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
|
||||
@@ -221,44 +221,91 @@ Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.094 | 5434.73 | 0.394 | 81.21 | 0.488 | 1114.15 |
|
||||
| 512 | 32 | 2 | 1088 | 0.168 | 6091.68 | 0.498 | 128.52 | 0.666 | 1633.41 |
|
||||
| 512 | 32 | 4 | 2176 | 0.341 | 6010.68 | 0.542 | 236.37 | 0.882 | 2466.43 |
|
||||
| 512 | 32 | 8 | 4352 | 0.665 | 6161.46 | 0.678 | 377.74 | 1.342 | 3241.72 |
|
||||
| 512 | 32 | 16 | 8704 | 1.323 | 6193.19 | 0.902 | 567.41 | 2.225 | 3911.74 |
|
||||
| 512 | 32 | 32 | 17408 | 2.642 | 6202.03 | 1.231 | 832.03 | 3.872 | 4495.36 |
|
||||
| 4096 | 32 | 1 | 4128 | 0.701 | 5840.49 | 0.439 | 72.95 | 1.140 | 3621.23 |
|
||||
| 4096 | 32 | 2 | 8256 | 1.387 | 5906.82 | 0.574 | 111.48 | 1.961 | 4210.12 |
|
||||
| 4096 | 32 | 4 | 16512 | 2.758 | 5940.33 | 0.651 | 196.58 | 3.409 | 4843.33 |
|
||||
| 4096 | 32 | 8 | 33024 | 5.491 | 5967.56 | 0.876 | 292.40 | 6.367 | 5187.12 |
|
||||
| 4096 | 32 | 16 | 66048 | 10.978 | 5969.58 | 1.275 | 401.69 | 12.253 | 5390.38 |
|
||||
| 4096 | 32 | 32 | 132096 | 21.944 | 5972.93 | 1.992 | 514.16 | 23.936 | 5518.73 |
|
||||
| 8192 | 32 | 1 | 8224 | 1.402 | 5841.91 | 0.452 | 70.73 | 1.855 | 4434.12 |
|
||||
| 8192 | 32 | 2 | 16448 | 2.793 | 5865.34 | 0.637 | 100.55 | 3.430 | 4795.51 |
|
||||
| 8192 | 32 | 4 | 32896 | 5.564 | 5889.64 | 0.770 | 166.26 | 6.334 | 5193.95 |
|
||||
| 8192 | 32 | 8 | 65792 | 11.114 | 5896.44 | 1.122 | 228.07 | 12.237 | 5376.51 |
|
||||
| 8192 | 32 | 16 | 131584 | 22.210 | 5901.38 | 1.789 | 286.15 | 24.000 | 5482.74 |
|
||||
| 8192 | 32 | 32 | 263168 | 44.382 | 5906.56 | 3.044 | 336.38 | 47.426 | 5549.02 |
|
||||
| 512 | 32 | 1 | 544 | 0.092 | 5566.97 | 0.412 | 77.63 | 0.504 | 1078.95 |
|
||||
| 512 | 32 | 2 | 1088 | 0.161 | 6345.67 | 0.522 | 122.70 | 0.683 | 1593.06 |
|
||||
| 512 | 32 | 4 | 2176 | 0.325 | 6309.87 | 0.562 | 227.68 | 0.887 | 2453.87 |
|
||||
| 512 | 32 | 8 | 4352 | 0.643 | 6374.42 | 0.685 | 373.67 | 1.328 | 3277.94 |
|
||||
| 512 | 32 | 16 | 8704 | 1.277 | 6413.64 | 0.915 | 559.47 | 2.192 | 3970.01 |
|
||||
| 512 | 32 | 32 | 17408 | 2.518 | 6506.57 | 1.249 | 819.61 | 3.767 | 4620.64 |
|
||||
| 4096 | 32 | 1 | 4128 | 0.674 | 6079.68 | 0.453 | 70.60 | 1.127 | 3662.88 |
|
||||
| 4096 | 32 | 2 | 8256 | 1.335 | 6137.82 | 0.627 | 102.03 | 1.962 | 4208.11 |
|
||||
| 4096 | 32 | 4 | 16512 | 2.657 | 6167.35 | 0.749 | 170.92 | 3.405 | 4848.71 |
|
||||
| 4096 | 32 | 8 | 33024 | 5.307 | 6173.91 | 0.974 | 262.89 | 6.281 | 5257.53 |
|
||||
| 4096 | 32 | 16 | 66048 | 10.610 | 6176.96 | 1.379 | 371.42 | 11.988 | 5509.40 |
|
||||
| 4096 | 32 | 32 | 132096 | 21.213 | 6178.89 | 2.122 | 482.50 | 23.335 | 5660.82 |
|
||||
| 8192 | 32 | 1 | 8224 | 1.359 | 6027.34 | 0.467 | 68.52 | 1.826 | 4503.48 |
|
||||
| 8192 | 32 | 2 | 16448 | 2.699 | 6069.68 | 0.653 | 98.03 | 3.352 | 4906.68 |
|
||||
| 8192 | 32 | 4 | 32896 | 5.366 | 6106.74 | 0.818 | 156.55 | 6.184 | 5319.96 |
|
||||
| 8192 | 32 | 8 | 65792 | 10.755 | 6093.50 | 1.174 | 218.04 | 11.929 | 5515.22 |
|
||||
| 8192 | 32 | 16 | 131584 | 21.484 | 6100.82 | 1.829 | 279.90 | 23.314 | 5644.11 |
|
||||
| 8192 | 32 | 32 | 263168 | 42.950 | 6103.40 | 3.058 | 334.91 | 46.008 | 5720.05 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 | 5810.04 ± 21.71 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 | 84.54 ± 0.18 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d4096 | 5288.04 ± 3.54 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d4096 | 78.82 ± 1.37 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d8192 | 4960.43 ± 16.64 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d8192 | 74.13 ± 0.30 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d16384 | 4495.92 ± 31.11 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d16384 | 72.37 ± 0.29 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | pp2048 @ d32768 | 3746.90 ± 40.01 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | tg32 @ d32768 | 63.02 ± 0.20 |
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | mmap | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 | 5948.74 ± 10.61 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 | 81.05 ± 0.20 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d4096 | 5652.69 ± 34.29 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d4096 | 76.37 ± 0.58 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d8192 | 5509.57 ± 40.69 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d8192 | 71.61 ± 0.80 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d16384 | 5340.86 ± 36.92 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d16384 | 70.89 ± 0.34 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | pp2048 @ d32768 | 5023.30 ± 13.52 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | CUDA | 99 | 2048 | 1 | 0 | 1 | tg32 @ d32768 | 62.28 ± 0.30 |
|
||||
|
||||
build: eeee367de (6989)
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
## ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.433 | 1181.83 | 0.693 | 46.16 | 1.126 | 482.94 |
|
||||
| 512 | 32 | 2 | 1088 | 0.439 | 2334.46 | 1.034 | 61.89 | 1.473 | 738.75 |
|
||||
| 512 | 32 | 4 | 2176 | 0.772 | 2654.46 | 1.459 | 87.76 | 2.230 | 975.77 |
|
||||
| 512 | 32 | 8 | 4352 | 1.541 | 2658.78 | 2.043 | 125.31 | 3.583 | 1214.47 |
|
||||
| 512 | 32 | 16 | 8704 | 3.083 | 2656.91 | 2.675 | 191.42 | 5.758 | 1511.62 |
|
||||
| 512 | 32 | 32 | 17408 | 6.159 | 2660.12 | 3.615 | 283.24 | 9.774 | 1780.98 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.915 | 2139.30 | 0.725 | 44.14 | 2.640 | 1563.83 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.834 | 2136.40 | 1.119 | 57.21 | 4.953 | 1666.81 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.636 | 2145.72 | 1.631 | 78.49 | 9.266 | 1781.93 |
|
||||
| 4096 | 32 | 8 | 33024 | 15.295 | 2142.40 | 2.344 | 109.21 | 17.639 | 1872.20 |
|
||||
| 4096 | 32 | 16 | 66048 | 30.573 | 2143.62 | 3.773 | 135.70 | 34.346 | 1923.04 |
|
||||
| 4096 | 32 | 32 | 132096 | 61.282 | 2138.82 | 5.795 | 176.71 | 67.077 | 1969.31 |
|
||||
| 8192 | 32 | 1 | 8224 | 4.510 | 1816.24 | 0.760 | 42.11 | 5.270 | 1560.44 |
|
||||
| 8192 | 32 | 2 | 16448 | 9.036 | 1813.19 | 1.206 | 53.06 | 10.242 | 1605.91 |
|
||||
| 8192 | 32 | 4 | 32896 | 18.070 | 1813.43 | 1.783 | 71.80 | 19.852 | 1657.03 |
|
||||
| 8192 | 32 | 8 | 65792 | 36.125 | 1814.15 | 2.635 | 97.14 | 38.760 | 1697.41 |
|
||||
| 8192 | 32 | 16 | 131584 | 72.367 | 1811.20 | 4.954 | 103.34 | 77.322 | 1701.77 |
|
||||
| 8192 | 32 | 32 | 263168 | 144.501 | 1814.13 | 8.103 | 126.37 | 152.604 | 1724.51 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | ngl | n_ubatch | fa | dio | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --: | --------------: | -------------------: |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 | 2364.18 ± 11.43 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 | 48.68 ± 0.12 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d4096 | 1684.13 ± 1.24 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d4096 | 44.62 ± 0.22 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d8192 | 1314.68 ± 1.41 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d8192 | 42.59 ± 0.11 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d16384 | 914.05 ± 3.32 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d16384 | 38.72 ± 0.13 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | pp2048 @ d32768 | 567.20 ± 0.90 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | CUDA | 99 | 2048 | 1 | 1 | tg32 @ d32768 | 32.65 ± 0.09 |
|
||||
|
||||
build: 11fb327bf (7941)
|
||||
|
||||
@@ -0,0 +1,298 @@
|
||||
## System info
|
||||
|
||||
```bash
|
||||
uname -a
|
||||
Darwin gg-studio 25.2.0 Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020 arm64
|
||||
|
||||
g++ --version
|
||||
Apple clang version 17.0.0 (clang-1700.3.19.1)
|
||||
Target: arm64-apple-darwin25.2.0
|
||||
```
|
||||
|
||||
## ggml-org/gpt-oss-20b-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.215 | 2381.35 | 0.245 | 130.45 | 0.460 | 1181.81 |
|
||||
| 512 | 32 | 2 | 1088 | 0.379 | 2701.43 | 0.382 | 167.56 | 0.761 | 1429.67 |
|
||||
| 512 | 32 | 4 | 2176 | 0.721 | 2839.27 | 0.604 | 211.76 | 1.326 | 1641.32 |
|
||||
| 512 | 32 | 8 | 4352 | 1.433 | 2858.30 | 1.033 | 247.75 | 2.466 | 1764.57 |
|
||||
| 512 | 32 | 16 | 8704 | 2.853 | 2871.12 | 1.570 | 326.11 | 4.423 | 1967.77 |
|
||||
| 512 | 32 | 32 | 17408 | 5.699 | 2874.95 | 1.910 | 536.15 | 7.609 | 2287.88 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.552 | 2638.56 | 0.334 | 95.72 | 1.887 | 2188.00 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.084 | 2655.88 | 0.404 | 158.54 | 3.488 | 2366.86 |
|
||||
| 4096 | 32 | 4 | 16512 | 6.151 | 2663.78 | 0.652 | 196.39 | 6.802 | 2427.37 |
|
||||
| 4096 | 32 | 8 | 33024 | 12.288 | 2666.77 | 1.135 | 225.47 | 13.423 | 2460.27 |
|
||||
| 4096 | 32 | 16 | 66048 | 24.563 | 2668.12 | 1.762 | 290.55 | 26.325 | 2508.97 |
|
||||
| 4096 | 32 | 32 | 132096 | 49.114 | 2668.73 | 2.398 | 426.94 | 51.512 | 2564.35 |
|
||||
| 8192 | 32 | 1 | 8224 | 3.345 | 2448.78 | 0.275 | 116.46 | 3.620 | 2271.76 |
|
||||
| 8192 | 32 | 2 | 16448 | 6.665 | 2458.11 | 0.425 | 150.71 | 7.090 | 2319.91 |
|
||||
| 8192 | 32 | 4 | 32896 | 13.315 | 2460.92 | 0.691 | 185.21 | 14.006 | 2348.63 |
|
||||
| 8192 | 32 | 8 | 65792 | 26.611 | 2462.73 | 1.212 | 211.16 | 27.823 | 2364.62 |
|
||||
| 8192 | 32 | 16 | 131584 | 53.232 | 2462.27 | 1.919 | 266.83 | 55.151 | 2385.88 |
|
||||
| 8192 | 32 | 32 | 263168 | 110.455 | 2373.30 | 2.752 | 372.03 | 113.208 | 2324.64 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 2713.40 ± 3.56 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 129.97 ± 3.90 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 2324.59 ± 3.01 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 123.38 ± 0.17 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1989.82 ± 30.11 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 117.39 ± 0.33 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 1556.54 ± 6.22 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 109.75 ± 0.42 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 1122.63 ± 1.45 |
|
||||
| gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 98.25 ± 0.08 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/gpt-oss-120b-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.426 | 1200.92 | 0.361 | 88.56 | 0.788 | 690.64 |
|
||||
| 512 | 32 | 2 | 1088 | 0.683 | 1500.14 | 0.545 | 117.35 | 1.228 | 886.02 |
|
||||
| 512 | 32 | 4 | 2176 | 1.204 | 1701.56 | 0.847 | 151.19 | 2.050 | 1061.34 |
|
||||
| 512 | 32 | 8 | 4352 | 2.402 | 1705.20 | 1.455 | 176.00 | 3.857 | 1128.45 |
|
||||
| 512 | 32 | 16 | 8704 | 4.802 | 1705.90 | 2.349 | 217.93 | 7.152 | 1217.08 |
|
||||
| 512 | 32 | 32 | 17408 | 9.593 | 1707.85 | 3.665 | 279.42 | 13.258 | 1313.01 |
|
||||
| 4096 | 32 | 1 | 4128 | 2.581 | 1587.08 | 0.390 | 82.12 | 2.970 | 1389.67 |
|
||||
| 4096 | 32 | 2 | 8256 | 5.124 | 1598.79 | 0.589 | 108.62 | 5.713 | 1445.10 |
|
||||
| 4096 | 32 | 4 | 16512 | 10.231 | 1601.47 | 0.928 | 137.98 | 11.158 | 1479.80 |
|
||||
| 4096 | 32 | 8 | 33024 | 20.468 | 1600.94 | 1.606 | 159.38 | 22.074 | 1496.04 |
|
||||
| 4096 | 32 | 16 | 66048 | 40.924 | 1601.42 | 2.639 | 193.99 | 43.563 | 1516.15 |
|
||||
| 4096 | 32 | 32 | 132096 | 81.819 | 1601.98 | 4.466 | 229.29 | 86.284 | 1530.94 |
|
||||
| 8192 | 32 | 1 | 8224 | 5.517 | 1484.74 | 0.409 | 78.16 | 5.927 | 1387.58 |
|
||||
| 8192 | 32 | 2 | 16448 | 11.008 | 1488.43 | 0.622 | 102.92 | 11.629 | 1414.34 |
|
||||
| 8192 | 32 | 4 | 32896 | 22.002 | 1489.29 | 0.987 | 129.66 | 22.990 | 1430.90 |
|
||||
| 8192 | 32 | 8 | 65792 | 46.051 | 1423.11 | 1.858 | 137.79 | 47.909 | 1373.27 |
|
||||
| 8192 | 32 | 16 | 131584 | 97.680 | 1341.85 | 2.872 | 178.28 | 100.552 | 1308.62 |
|
||||
| 8192 | 32 | 32 | 263168 | 176.407 | 1486.02 | 5.048 | 202.85 | 181.455 | 1450.32 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 1648.69 ± 1.80 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 85.60 ± 0.52 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 1429.86 ± 1.01 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 82.03 ± 0.12 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1257.90 ± 1.81 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 78.23 ± 0.33 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 1013.49 ± 0.70 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 73.20 ± 0.28 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 721.11 ± 0.58 |
|
||||
| gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 65.52 ± 0.10 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.243 | 2109.23 | 0.419 | 76.34 | 0.662 | 821.84 |
|
||||
| 512 | 32 | 2 | 1088 | 0.406 | 2521.40 | 0.575 | 111.36 | 0.981 | 1109.27 |
|
||||
| 512 | 32 | 4 | 2176 | 0.744 | 2751.65 | 0.841 | 152.22 | 1.585 | 1372.71 |
|
||||
| 512 | 32 | 8 | 4352 | 1.479 | 2770.20 | 1.330 | 192.48 | 2.809 | 1549.53 |
|
||||
| 512 | 32 | 16 | 8704 | 2.951 | 2776.20 | 2.572 | 199.05 | 5.523 | 1575.93 |
|
||||
| 512 | 32 | 32 | 17408 | 5.899 | 2777.64 | 2.603 | 393.34 | 8.502 | 2047.54 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.901 | 2154.15 | 0.474 | 67.58 | 2.375 | 1738.14 |
|
||||
| 4096 | 32 | 2 | 8256 | 3.788 | 2162.89 | 0.652 | 98.17 | 4.439 | 1859.69 |
|
||||
| 4096 | 32 | 4 | 16512 | 7.564 | 2166.18 | 0.990 | 129.24 | 8.554 | 1930.34 |
|
||||
| 4096 | 32 | 8 | 33024 | 15.121 | 2166.98 | 1.632 | 156.82 | 16.754 | 1971.12 |
|
||||
| 4096 | 32 | 16 | 66048 | 30.241 | 2167.09 | 3.166 | 161.72 | 33.407 | 1977.04 |
|
||||
| 4096 | 32 | 32 | 132096 | 60.474 | 2167.42 | 3.780 | 270.93 | 64.254 | 2055.86 |
|
||||
| 8192 | 32 | 1 | 8224 | 4.733 | 1730.92 | 0.483 | 66.29 | 5.215 | 1576.85 |
|
||||
| 8192 | 32 | 2 | 16448 | 9.459 | 1732.09 | 0.722 | 88.58 | 10.182 | 1615.46 |
|
||||
| 8192 | 32 | 4 | 32896 | 18.912 | 1732.65 | 1.120 | 114.26 | 20.032 | 1642.14 |
|
||||
| 8192 | 32 | 8 | 65792 | 37.797 | 1733.91 | 1.873 | 136.67 | 39.670 | 1658.49 |
|
||||
| 8192 | 32 | 16 | 131584 | 84.133 | 1557.92 | 3.718 | 137.72 | 87.850 | 1497.82 |
|
||||
| 8192 | 32 | 32 | 263168 | 157.550 | 1663.88 | 4.854 | 210.98 | 162.403 | 1620.46 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 2453.11 ± 1.70 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 78.97 ± 0.46 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 1569.46 ± 1.97 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 71.18 ± 0.37 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1145.51 ± 1.16 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 65.11 ± 0.36 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 741.04 ± 0.74 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 56.87 ± 0.14 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 431.31 ± 0.31 |
|
||||
| qwen3moe 30B.A3B Q8_0 | 30.25 GiB | 30.53 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 45.26 ± 0.11 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.339 | 1509.22 | 0.409 | 78.17 | 0.749 | 726.67 |
|
||||
| 512 | 32 | 2 | 1088 | 0.646 | 1584.93 | 0.483 | 132.45 | 1.129 | 963.45 |
|
||||
| 512 | 32 | 4 | 2176 | 1.258 | 1627.50 | 0.585 | 218.67 | 1.844 | 1180.21 |
|
||||
| 512 | 32 | 8 | 4352 | 2.506 | 1634.41 | 1.005 | 254.83 | 3.511 | 1239.64 |
|
||||
| 512 | 32 | 16 | 8704 | 5.007 | 1635.99 | 1.595 | 321.07 | 6.602 | 1318.38 |
|
||||
| 512 | 32 | 32 | 17408 | 10.007 | 1637.19 | 1.676 | 611.12 | 11.683 | 1490.03 |
|
||||
| 4096 | 32 | 1 | 4128 | 2.730 | 1500.46 | 0.431 | 74.31 | 3.160 | 1306.12 |
|
||||
| 4096 | 32 | 2 | 8256 | 5.446 | 1504.33 | 0.524 | 122.04 | 5.970 | 1382.91 |
|
||||
| 4096 | 32 | 4 | 16512 | 10.875 | 1506.59 | 0.662 | 193.45 | 11.537 | 1431.28 |
|
||||
| 4096 | 32 | 8 | 33024 | 21.749 | 1506.61 | 1.158 | 221.11 | 22.907 | 1441.64 |
|
||||
| 4096 | 32 | 16 | 66048 | 43.477 | 1507.36 | 1.901 | 269.32 | 45.378 | 1455.49 |
|
||||
| 4096 | 32 | 32 | 132096 | 86.954 | 1507.37 | 2.325 | 440.42 | 89.279 | 1479.59 |
|
||||
| 8192 | 32 | 1 | 8224 | 5.940 | 1379.21 | 0.449 | 71.20 | 6.389 | 1287.20 |
|
||||
| 8192 | 32 | 2 | 16448 | 11.865 | 1380.84 | 0.559 | 114.59 | 12.424 | 1323.92 |
|
||||
| 8192 | 32 | 4 | 32896 | 23.723 | 1381.25 | 0.728 | 175.80 | 24.452 | 1345.35 |
|
||||
| 8192 | 32 | 8 | 65792 | 47.434 | 1381.63 | 1.279 | 200.09 | 48.713 | 1350.60 |
|
||||
| 8192 | 32 | 16 | 131584 | 94.864 | 1381.69 | 2.198 | 232.97 | 97.061 | 1355.68 |
|
||||
| 8192 | 32 | 32 | 263168 | 189.743 | 1381.57 | 3.052 | 335.50 | 192.795 | 1365.01 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 1565.91 ± 0.86 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 79.68 ± 0.39 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 1317.41 ± 1.02 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 74.70 ± 0.04 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 1134.65 ± 0.76 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 71.31 ± 0.12 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 886.46 ± 0.78 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 65.93 ± 0.06 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 612.21 ± 0.30 |
|
||||
| qwen2 7B Q8_0 | 7.54 GiB | 7.62 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 56.83 ± 0.02 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.186 | 2748.06 | 0.235 | 136.28 | 0.421 | 1291.78 |
|
||||
| 512 | 32 | 2 | 1088 | 0.342 | 2990.95 | 0.312 | 204.99 | 0.655 | 1662.15 |
|
||||
| 512 | 32 | 4 | 2176 | 0.662 | 3092.69 | 0.404 | 316.97 | 1.066 | 2041.21 |
|
||||
| 512 | 32 | 8 | 4352 | 1.317 | 3110.41 | 0.579 | 441.80 | 1.896 | 2294.97 |
|
||||
| 512 | 32 | 16 | 8704 | 2.625 | 3120.23 | 1.207 | 424.08 | 3.833 | 2270.93 |
|
||||
| 512 | 32 | 32 | 17408 | 5.242 | 3125.34 | 1.299 | 788.23 | 6.541 | 2661.19 |
|
||||
| 4096 | 32 | 1 | 4128 | 1.408 | 2909.90 | 0.296 | 108.07 | 1.704 | 2422.95 |
|
||||
| 4096 | 32 | 2 | 8256 | 2.793 | 2933.40 | 0.325 | 197.00 | 3.118 | 2648.25 |
|
||||
| 4096 | 32 | 4 | 16512 | 5.567 | 2943.22 | 0.440 | 291.07 | 6.006 | 2749.05 |
|
||||
| 4096 | 32 | 8 | 33024 | 11.114 | 2948.23 | 0.640 | 400.26 | 11.754 | 2809.59 |
|
||||
| 4096 | 32 | 16 | 66048 | 22.217 | 2949.76 | 1.327 | 385.83 | 23.544 | 2805.26 |
|
||||
| 4096 | 32 | 32 | 132096 | 44.420 | 2950.77 | 1.553 | 659.30 | 45.973 | 2873.36 |
|
||||
| 8192 | 32 | 1 | 8224 | 2.860 | 2864.58 | 0.250 | 127.90 | 3.110 | 2644.42 |
|
||||
| 8192 | 32 | 2 | 16448 | 5.702 | 2873.63 | 0.335 | 191.07 | 6.036 | 2724.77 |
|
||||
| 8192 | 32 | 4 | 32896 | 11.383 | 2878.69 | 0.456 | 280.72 | 11.839 | 2778.63 |
|
||||
| 8192 | 32 | 8 | 65792 | 22.750 | 2880.75 | 0.671 | 381.48 | 23.421 | 2809.14 |
|
||||
| 8192 | 32 | 16 | 131584 | 45.484 | 2881.74 | 1.406 | 364.04 | 46.890 | 2806.22 |
|
||||
| 8192 | 32 | 32 | 263168 | 90.956 | 2882.10 | 1.793 | 570.98 | 92.749 | 2837.41 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 2923.59 ± 3.10 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 134.28 ± 1.29 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 2748.21 ± 3.05 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 133.11 ± 0.08 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 2641.45 ± 2.31 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 125.85 ± 0.35 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 2446.20 ± 2.94 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 125.00 ± 0.12 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 2129.18 ± 7.43 |
|
||||
| gemma3 4B Q4_0 | 2.35 GiB | 3.88 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 113.14 ± 0.10 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
|
||||
## ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
|
||||
|
||||
- `llama-batched-bench`
|
||||
|
||||
|
||||
main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
|
||||
|
||||
| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s |
|
||||
|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
|
||||
| 512 | 32 | 1 | 544 | 0.326 | 1568.69 | 0.522 | 61.28 | 0.849 | 641.09 |
|
||||
| 512 | 32 | 2 | 1088 | 0.528 | 1939.42 | 0.744 | 86.07 | 1.272 | 855.63 |
|
||||
| 512 | 32 | 4 | 2176 | 0.968 | 2114.85 | 1.105 | 115.85 | 2.073 | 1049.56 |
|
||||
| 512 | 32 | 8 | 4352 | 1.928 | 2124.62 | 1.684 | 151.99 | 3.612 | 1204.82 |
|
||||
| 512 | 32 | 16 | 8704 | 3.844 | 2131.34 | 3.141 | 162.99 | 6.985 | 1246.11 |
|
||||
| 512 | 32 | 32 | 17408 | 7.683 | 2132.38 | 3.924 | 260.95 | 11.608 | 1499.71 |
|
||||
| 4096 | 32 | 1 | 4128 | 3.280 | 1248.75 | 0.723 | 44.29 | 4.003 | 1031.33 |
|
||||
| 4096 | 32 | 2 | 8256 | 6.545 | 1251.63 | 0.930 | 68.85 | 7.475 | 1104.53 |
|
||||
| 4096 | 32 | 4 | 16512 | 13.080 | 1252.64 | 1.454 | 88.03 | 14.534 | 1136.12 |
|
||||
| 4096 | 32 | 8 | 33024 | 26.154 | 1252.90 | 2.388 | 107.20 | 28.542 | 1157.04 |
|
||||
| 4096 | 32 | 16 | 66048 | 52.297 | 1253.14 | 4.724 | 108.37 | 57.022 | 1158.30 |
|
||||
| 4096 | 32 | 32 | 132096 | 104.578 | 1253.34 | 7.266 | 140.93 | 111.844 | 1181.08 |
|
||||
| 8192 | 32 | 1 | 8224 | 9.623 | 851.31 | 0.767 | 41.72 | 10.390 | 791.54 |
|
||||
| 8192 | 32 | 2 | 16448 | 20.916 | 783.32 | 1.148 | 55.74 | 22.064 | 745.45 |
|
||||
| 8192 | 32 | 4 | 32896 | 43.509 | 753.14 | 1.833 | 69.82 | 45.342 | 725.51 |
|
||||
| 8192 | 32 | 8 | 65792 | 79.621 | 823.10 | 3.180 | 80.50 | 82.801 | 794.58 |
|
||||
| 8192 | 32 | 16 | 131584 | 153.770 | 852.39 | 6.502 | 78.74 | 160.272 | 821.00 |
|
||||
| 8192 | 32 | 32 | 263168 | 307.539 | 852.39 | 10.839 | 94.48 | 318.378 | 826.59 |
|
||||
|
||||
|
||||
- `llama-bench`
|
||||
|
||||
| model | size | params | backend | threads | n_ubatch | fa | test | t/s |
|
||||
| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 | 1629.33 ± 0.27 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 | 59.58 ± 0.13 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d4096 | 732.67 ± 0.42 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d4096 | 47.44 ± 0.15 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d8192 | 474.33 ± 0.33 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d8192 | 40.20 ± 0.20 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d16384 | 277.46 ± 0.09 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d16384 | 31.50 ± 0.93 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | pp2048 @ d32768 | 151.44 ± 0.05 |
|
||||
| deepseek2 30B.A3B Q8_0 | 29.65 GiB | 29.94 B | MTL,BLAS | 16 | 2048 | 1 | tg32 @ d32768 | 21.81 ± 0.01 |
|
||||
|
||||
build: b828e18c7 (7948)
|
||||
@@ -413,8 +413,6 @@ function gg_run_qwen3_0_6b {
|
||||
./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k $(nproc)
|
||||
./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k $(nproc)
|
||||
|
||||
(time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
|
||||
|
||||
(time ./bin/llama-completion -no-cnv --model ${model_f16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
|
||||
(time ./bin/llama-completion -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
|
||||
(time ./bin/llama-completion -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
|
||||
@@ -540,8 +538,6 @@ function gg_run_embd_bge_small {
|
||||
|
||||
./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
|
||||
|
||||
(time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
|
||||
|
||||
(time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
|
||||
(time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
|
||||
|
||||
@@ -582,8 +578,6 @@ function gg_run_rerank_tiny {
|
||||
|
||||
model_f16="${path_models}/ggml-model-f16.gguf"
|
||||
|
||||
(time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
|
||||
|
||||
# for this model, the SEP token is "</s>"
|
||||
(time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
|
||||
|
||||
@@ -683,8 +677,8 @@ fi
|
||||
|
||||
ret=0
|
||||
|
||||
test $ret -eq 0 && gg_run ctest_debug
|
||||
test $ret -eq 0 && gg_run ctest_release
|
||||
#test $ret -eq 0 && gg_run ctest_debug
|
||||
#test $ret -eq 0 && gg_run ctest_release
|
||||
|
||||
if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
|
||||
test $ret -eq 0 && gg_run test_backend_ops_cpu
|
||||
|
||||
@@ -32,4 +32,27 @@ function(llama_add_compile_flags)
|
||||
set(CXX_FLAGS "" PARENT_SCOPE)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if (NOT MSVC)
|
||||
if (LLAMA_SANITIZE_THREAD)
|
||||
message(STATUS "Using -fsanitize=thread")
|
||||
|
||||
add_compile_options(-fsanitize=thread)
|
||||
link_libraries (-fsanitize=thread)
|
||||
endif()
|
||||
|
||||
if (LLAMA_SANITIZE_ADDRESS)
|
||||
message(STATUS "Using -fsanitize=address")
|
||||
|
||||
add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
|
||||
link_libraries (-fsanitize=address)
|
||||
endif()
|
||||
|
||||
if (LLAMA_SANITIZE_UNDEFINED)
|
||||
message(STATUS "Using -fsanitize=undefined")
|
||||
|
||||
add_compile_options(-fsanitize=undefined)
|
||||
link_libraries (-fsanitize=undefined)
|
||||
endif()
|
||||
endif()
|
||||
endfunction()
|
||||
|
||||
+9
-7
@@ -2331,19 +2331,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
|
||||
}
|
||||
).set_env("LLAMA_ARG_N_GPU_LAYERS"));
|
||||
add_opt(common_arg(
|
||||
{"-sm", "--split-mode"}, "{none,layer,row}",
|
||||
{"-sm", "--split-mode"}, "{none,layer,row,tensor}",
|
||||
"how to split the model across multiple GPUs, one of:\n"
|
||||
"- none: use one GPU only\n"
|
||||
"- layer (default): split layers and KV across GPUs\n"
|
||||
"- row: split rows across GPUs",
|
||||
"- layer (default): split layers and KV across GPUs (pipelined)\n"
|
||||
"- row: split weight across GPUs by rows (parallelized)\n"
|
||||
"- tensor: split weights and KV across GPUs (parallelized)",
|
||||
[](common_params & params, const std::string & value) {
|
||||
std::string arg_next = value;
|
||||
if (arg_next == "none") {
|
||||
if (value == "none") {
|
||||
params.split_mode = LLAMA_SPLIT_MODE_NONE;
|
||||
} else if (arg_next == "layer") {
|
||||
} else if (value == "layer") {
|
||||
params.split_mode = LLAMA_SPLIT_MODE_LAYER;
|
||||
} else if (arg_next == "row") {
|
||||
} else if (value == "row") {
|
||||
params.split_mode = LLAMA_SPLIT_MODE_ROW;
|
||||
} else if (value == "tensor") {
|
||||
params.split_mode = LLAMA_SPLIT_MODE_TENSOR;
|
||||
} else {
|
||||
throw std::invalid_argument("invalid value");
|
||||
}
|
||||
|
||||
+18
-16
@@ -45,6 +45,8 @@ static float common_ggml_get_float_value(const uint8_t * data,
|
||||
return v;
|
||||
}
|
||||
|
||||
#define INDENT " "
|
||||
|
||||
template <bool abort>
|
||||
void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
|
||||
GGML_ASSERT(n > 0);
|
||||
@@ -60,41 +62,41 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
|
||||
}
|
||||
}
|
||||
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
|
||||
LOG_ERR(" [\n");
|
||||
LOG(INDENT "[\n");
|
||||
for (int64_t i2 = 0; i2 < ne[2]; i2++) {
|
||||
if (i2 == n && ne[2] > 2 * n) {
|
||||
LOG_ERR(" ..., \n");
|
||||
LOG(INDENT INDENT "..., \n");
|
||||
i2 = ne[2] - n;
|
||||
}
|
||||
LOG_ERR(" [\n");
|
||||
LOG(INDENT INDENT "[\n");
|
||||
for (int64_t i1 = 0; i1 < ne[1]; i1++) {
|
||||
if (i1 == n && ne[1] > 2 * n) {
|
||||
LOG_ERR(" ..., \n");
|
||||
LOG(INDENT INDENT INDENT "..., \n");
|
||||
i1 = ne[1] - n;
|
||||
}
|
||||
LOG_ERR(" [");
|
||||
LOG(INDENT INDENT INDENT "[");
|
||||
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
|
||||
if (i0 == n && ne[0] > 2 * n) {
|
||||
LOG_ERR("..., ");
|
||||
LOG(" ..., ");
|
||||
i0 = ne[0] - n;
|
||||
}
|
||||
const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
|
||||
LOG_ERR("%12.4f", v);
|
||||
LOG("%12.4f", v);
|
||||
if (i0 < ne[0] - 1) {
|
||||
LOG_ERR(", ");
|
||||
LOG(", ");
|
||||
}
|
||||
}
|
||||
LOG_ERR("],\n");
|
||||
LOG(" ],\n");
|
||||
}
|
||||
LOG_ERR(" ],\n");
|
||||
LOG(INDENT INDENT "],\n");
|
||||
}
|
||||
LOG_ERR(" ]\n");
|
||||
LOG_ERR(" sum = %f\n", sum);
|
||||
LOG(INDENT "]\n");
|
||||
LOG(INDENT "sum = %f\n", sum);
|
||||
}
|
||||
|
||||
if constexpr (abort) {
|
||||
if (std::isnan(sum)) {
|
||||
LOG_ERR("encountered NaN - aborting\n");
|
||||
LOG("encountered NaN - aborting\n");
|
||||
exit(0);
|
||||
}
|
||||
}
|
||||
@@ -137,9 +139,9 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
|
||||
}
|
||||
|
||||
if (matches_filter) {
|
||||
LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
|
||||
ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
|
||||
common_ggml_ne_string(t).c_str());
|
||||
LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
|
||||
ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
|
||||
common_ggml_ne_string(t).c_str());
|
||||
}
|
||||
|
||||
const bool is_host = ggml_backend_buffer_is_host(t->buffer);
|
||||
|
||||
+3
-12
@@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start,
|
||||
* @return Vector of draft tokens, empty if no matching pattern is found
|
||||
*/
|
||||
llama_tokens common_ngram_simple_draft(
|
||||
common_ngram_simple_state & state,
|
||||
const common_ngram_simple_config & config,
|
||||
const llama_tokens & tokens, llama_token sampled) {
|
||||
|
||||
// Simple implementation of self-speculative decoding without a draft model.
|
||||
//
|
||||
const size_t cur_len = tokens.size();
|
||||
// Only check every check_rate tokens to save compute
|
||||
// i.e., perform check if (cur_len - idx_last_check) >= check_rate
|
||||
if (state.idx_last_check + state.config.check_rate > cur_len) {
|
||||
llama_tokens draft_tokens;
|
||||
return draft_tokens;
|
||||
}
|
||||
|
||||
size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
|
||||
size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
|
||||
const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history
|
||||
const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft
|
||||
|
||||
// vector for tokens we want to verify.
|
||||
// return empty vector if there is no match.
|
||||
@@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft(
|
||||
}
|
||||
pattern.push_back(sampled); // add the last token to the pattern
|
||||
|
||||
// We do a search in the token history.
|
||||
state.idx_last_check = cur_len;
|
||||
|
||||
size_t match_pos = 0; // we ignore position 0, position 0 == no match
|
||||
// search backwards, but skip the current match (we are currently there)
|
||||
for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {
|
||||
|
||||
+1
-15
@@ -27,23 +27,9 @@ struct common_ngram_simple_config {
|
||||
uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
|
||||
};
|
||||
|
||||
// current state (and config) of n-gram simple.
|
||||
struct common_ngram_simple_state {
|
||||
common_ngram_simple_config config;
|
||||
|
||||
size_t idx_last_check = 0; // index of last check in context history (mutable)
|
||||
|
||||
common_ngram_simple_state(const common_ngram_simple_config & config)
|
||||
: config(config) {}
|
||||
};
|
||||
|
||||
// Searches for a n-gram in the history and checks whether a draft sequence should be generated.
|
||||
// state: the ngram simple state to search in.
|
||||
// inp: the tokens generated so far.
|
||||
// sampled: the token that was just sampled.
|
||||
// draft: vector to store the draft tokens, initially empty.
|
||||
llama_tokens common_ngram_simple_draft(
|
||||
common_ngram_simple_state & state,
|
||||
const common_ngram_simple_config & config,
|
||||
const llama_tokens & tokens, llama_token sampled);
|
||||
|
||||
|
||||
|
||||
+14
-6
@@ -463,12 +463,14 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
|
||||
|
||||
// state of self-speculation (simple implementation, not ngram-map)
|
||||
struct common_speculative_state_ngram_simple : public common_speculative_state {
|
||||
common_ngram_simple_state state;
|
||||
common_ngram_simple_config config;
|
||||
|
||||
uint16_t check_id = 0; // used to control the frequency of generating drafts
|
||||
|
||||
common_speculative_state_ngram_simple(
|
||||
enum common_speculative_type type,
|
||||
common_ngram_simple_state state)
|
||||
: common_speculative_state(type), state(state) {}
|
||||
common_ngram_simple_config config)
|
||||
: common_speculative_state(type), config(config) {}
|
||||
|
||||
void begin(const llama_tokens & prompt) override {
|
||||
GGML_UNUSED(prompt);
|
||||
@@ -479,7 +481,13 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
|
||||
const llama_tokens & prompt_tgt,
|
||||
llama_token id_last,
|
||||
llama_tokens & result) override {
|
||||
result = common_ngram_simple_draft(state, prompt_tgt, id_last);
|
||||
++check_id;
|
||||
if (check_id < config.check_rate) {
|
||||
return;
|
||||
}
|
||||
check_id = 0;
|
||||
|
||||
result = common_ngram_simple_draft(config, prompt_tgt, id_last);
|
||||
GGML_UNUSED(params);
|
||||
}
|
||||
|
||||
@@ -889,14 +897,14 @@ common_speculative * common_speculative_init(
|
||||
uint16_t mgram_size_value = ngram_map.size_value;
|
||||
uint16_t check_rate = ngram_map.check_rate;
|
||||
|
||||
auto config_simple = common_ngram_simple_config{
|
||||
auto config_simple = common_ngram_simple_config {
|
||||
/* .size_ngram = */ ngram_size_key,
|
||||
/* .size_mgram = */ mgram_size_value,
|
||||
/* .check_rate = */ check_rate
|
||||
};
|
||||
auto state = std::make_unique<common_speculative_state_ngram_simple>(
|
||||
/* .type = */ config.type,
|
||||
/* .state = */ common_ngram_simple_state(config_simple)
|
||||
/* .state = */ config_simple
|
||||
);
|
||||
impls.push_back(std::move(state));
|
||||
break;
|
||||
|
||||
+159
@@ -0,0 +1,159 @@
|
||||
#!/usr/bin/env python3
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
from safetensors import safe_open
|
||||
|
||||
|
||||
MODEL_SAFETENSORS_FILE = "model.safetensors"
|
||||
MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
|
||||
|
||||
|
||||
def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
|
||||
index_file = model_path / MODEL_SAFETENSORS_INDEX
|
||||
|
||||
if index_file.exists():
|
||||
with open(index_file, 'r') as f:
|
||||
index = json.load(f)
|
||||
return index.get("weight_map", {})
|
||||
|
||||
return None
|
||||
|
||||
|
||||
def get_all_tensor_names(model_path: Path) -> list[str]:
|
||||
weight_map = get_weight_map(model_path)
|
||||
|
||||
if weight_map is not None:
|
||||
return list(weight_map.keys())
|
||||
|
||||
single_file = model_path / MODEL_SAFETENSORS_FILE
|
||||
if single_file.exists():
|
||||
try:
|
||||
with safe_open(single_file, framework="pt", device="cpu") as f:
|
||||
return list(f.keys())
|
||||
except Exception as e:
|
||||
print(f"Error reading {single_file}: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
print(f"Error: No safetensors files found in {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
|
||||
def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
|
||||
weight_map = get_weight_map(model_path)
|
||||
|
||||
if weight_map is not None:
|
||||
return weight_map.get(tensor_name)
|
||||
|
||||
single_file = model_path / MODEL_SAFETENSORS_FILE
|
||||
if single_file.exists():
|
||||
return single_file.name
|
||||
|
||||
return None
|
||||
|
||||
|
||||
def normalize_tensor_name(tensor_name: str) -> str:
|
||||
normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
|
||||
normalized = re.sub(r'\.\d+$', '.#', normalized)
|
||||
return normalized
|
||||
|
||||
|
||||
def list_all_tensors(model_path: Path, unique: bool = False):
|
||||
tensor_names = get_all_tensor_names(model_path)
|
||||
|
||||
if unique:
|
||||
seen = set()
|
||||
for tensor_name in sorted(tensor_names):
|
||||
normalized = normalize_tensor_name(tensor_name)
|
||||
if normalized not in seen:
|
||||
seen.add(normalized)
|
||||
print(normalized)
|
||||
else:
|
||||
for tensor_name in sorted(tensor_names):
|
||||
print(tensor_name)
|
||||
|
||||
|
||||
def print_tensor_info(model_path: Path, tensor_name: str):
|
||||
tensor_file = find_tensor_file(model_path, tensor_name)
|
||||
|
||||
if tensor_file is None:
|
||||
print(f"Error: Could not find tensor '{tensor_name}' in model index")
|
||||
print(f"Model path: {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
file_path = model_path / tensor_file
|
||||
|
||||
try:
|
||||
with safe_open(file_path, framework="pt", device="cpu") as f:
|
||||
if tensor_name in f.keys():
|
||||
tensor_slice = f.get_slice(tensor_name)
|
||||
shape = tensor_slice.get_shape()
|
||||
print(f"Tensor: {tensor_name}")
|
||||
print(f"File: {tensor_file}")
|
||||
print(f"Shape: {shape}")
|
||||
else:
|
||||
print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
|
||||
sys.exit(1)
|
||||
|
||||
except FileNotFoundError:
|
||||
print(f"Error: The file '{file_path}' was not found.")
|
||||
sys.exit(1)
|
||||
except Exception as e:
|
||||
print(f"An error occurred: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Print tensor information from a safetensors model"
|
||||
)
|
||||
parser.add_argument(
|
||||
"tensor_name",
|
||||
nargs="?", # optional (if --list is used for example)
|
||||
help="Name of the tensor to inspect"
|
||||
)
|
||||
parser.add_argument(
|
||||
"-m", "--model-path",
|
||||
type=Path,
|
||||
help="Path to the model directory (default: MODEL_PATH environment variable)"
|
||||
)
|
||||
parser.add_argument(
|
||||
"-l", "--list",
|
||||
action="store_true",
|
||||
help="List unique tensor patterns in the model (layer numbers replaced with #)"
|
||||
)
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
model_path = args.model_path
|
||||
if model_path is None:
|
||||
model_path_str = os.environ.get("MODEL_PATH")
|
||||
if model_path_str is None:
|
||||
print("Error: --model-path not provided and MODEL_PATH environment variable not set")
|
||||
sys.exit(1)
|
||||
model_path = Path(model_path_str)
|
||||
|
||||
if not model_path.exists():
|
||||
print(f"Error: Model path does not exist: {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
if not model_path.is_dir():
|
||||
print(f"Error: Model path is not a directory: {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
if args.list:
|
||||
list_all_tensors(model_path, unique=True)
|
||||
else:
|
||||
if args.tensor_name is None:
|
||||
print("Error: tensor_name is required when not using --list")
|
||||
sys.exit(1)
|
||||
print_tensor_info(model_path, args.tensor_name)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -68,7 +68,7 @@ extern "C" {
|
||||
GGML_API void ggml_backend_buffer_reset (ggml_backend_buffer_t buffer);
|
||||
|
||||
// tensor copy between different backends
|
||||
GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
|
||||
GGML_API void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor * dst);
|
||||
|
||||
//
|
||||
// Backend (stream)
|
||||
@@ -109,7 +109,18 @@ extern "C" {
|
||||
// the copy is performed after all the currently queued operations in backend_src
|
||||
// backend_dst will wait for the copy to complete before performing other operations
|
||||
// automatic fallback to sync copy if async is not supported
|
||||
GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
|
||||
GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
|
||||
|
||||
// asynchronous tensor shuffle
|
||||
// - src1, dst1 belong to backend_1
|
||||
// - src2, dst2 belong to backend_2
|
||||
// - src1 is copied to dst2
|
||||
// - src2 is copied to dst1
|
||||
// - both backends wait until both copies have completed
|
||||
GGML_API void ggml_backend_tensor_shfl_async(
|
||||
ggml_backend_t backend_1, ggml_backend_t backend_2,
|
||||
const struct ggml_tensor * src1, const struct ggml_tensor * src2,
|
||||
struct ggml_tensor * dst1, struct ggml_tensor * dst2);
|
||||
|
||||
GGML_API ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend);
|
||||
|
||||
@@ -135,7 +146,9 @@ extern "C" {
|
||||
// integrated GPU device using host memory
|
||||
GGML_BACKEND_DEVICE_TYPE_IGPU,
|
||||
// accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
|
||||
GGML_BACKEND_DEVICE_TYPE_ACCEL
|
||||
GGML_BACKEND_DEVICE_TYPE_ACCEL,
|
||||
// "meta" device wrapping multiple other devices for tensor parallelism
|
||||
GGML_BACKEND_DEVICE_TYPE_META,
|
||||
};
|
||||
|
||||
// functionality supported by the device
|
||||
@@ -211,6 +224,52 @@ extern "C" {
|
||||
};
|
||||
typedef struct ggml_backend_feature * (*ggml_backend_get_features_t)(ggml_backend_reg_t reg);
|
||||
|
||||
//
|
||||
// Meta backend
|
||||
//
|
||||
|
||||
enum ggml_backend_meta_split_state {
|
||||
// tensor split by tensor dimensions:
|
||||
GGML_BACKEND_SPLIT_STATE_BY_NE0 = 0,
|
||||
GGML_BACKEND_SPLIT_STATE_BY_NE1 = 1,
|
||||
GGML_BACKEND_SPLIT_STATE_BY_NE2 = 2,
|
||||
GGML_BACKEND_SPLIT_STATE_BY_NE3 = 3,
|
||||
|
||||
GGML_BACKEND_SPLIT_STATE_MIRRORED = 10, // all values on all backends
|
||||
GGML_BACKEND_SPLIT_STATE_PARTIAL = 11, // each backend has a partial sum
|
||||
|
||||
// for internal bookkeeping only:
|
||||
GGML_BACKEND_SPLIT_STATE_NONE = 98,
|
||||
GGML_BACKEND_SPLIT_STATE_UNKNOWN = 99,
|
||||
};
|
||||
|
||||
// function to assign split states for statically allocated tensors, compute tensor split states will be assigned to be compatible:
|
||||
typedef enum ggml_backend_meta_split_state (*ggml_backend_meta_get_split_state_t)(const struct ggml_tensor * tensor, void * userdata);
|
||||
|
||||
|
||||
GGML_API bool ggml_backend_dev_is_meta(ggml_backend_dev_t dev);
|
||||
GGML_API size_t ggml_backend_meta_dev_n_devs(ggml_backend_dev_t meta_dev);
|
||||
GGML_API ggml_backend_dev_t ggml_backend_meta_dev_simple_dev(ggml_backend_dev_t meta_dev, size_t index);
|
||||
|
||||
// create a new meta device from "simple" devices, meta buffer type/buffer/backend is then derived from this:
|
||||
GGML_API ggml_backend_dev_t ggml_backend_meta_device(
|
||||
ggml_backend_dev_t * devs, size_t n_devs, ggml_backend_meta_get_split_state_t get_split_state, void * get_split_state_ud);
|
||||
|
||||
GGML_API bool ggml_backend_buft_is_meta(ggml_backend_buffer_type_t buft);
|
||||
GGML_API size_t ggml_backend_meta_buft_n_bufts(ggml_backend_buffer_type_t meta_buft);
|
||||
GGML_API ggml_backend_buffer_type_t ggml_backend_meta_buft_simple_buft(ggml_backend_buffer_type_t meta_buft, size_t index);
|
||||
|
||||
GGML_API bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf);
|
||||
GGML_API size_t ggml_backend_meta_buffer_n_bufs(ggml_backend_buffer_t meta_buf);
|
||||
GGML_API ggml_backend_buffer_t ggml_backend_meta_buffer_simple_buffer(ggml_backend_buffer_t meta_buf, size_t index);
|
||||
GGML_API struct ggml_tensor * ggml_backend_meta_buffer_simple_tensor(const struct ggml_tensor * tensor, size_t index);
|
||||
|
||||
GGML_API bool ggml_backend_is_meta(ggml_backend_t backend);
|
||||
GGML_API size_t ggml_backend_meta_n_backends(ggml_backend_t meta_backend);
|
||||
GGML_API ggml_backend_t ggml_backend_meta_simple_backend(ggml_backend_t meta_backend, size_t index);
|
||||
|
||||
GGML_API enum ggml_backend_meta_split_state ggml_backend_meta_get_split_state(const struct ggml_tensor * tensor, bool assume_sync);
|
||||
|
||||
//
|
||||
// Backend registry
|
||||
//
|
||||
|
||||
@@ -7,8 +7,6 @@
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend"
|
||||
|
||||
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg();
|
||||
|
||||
#ifdef __cplusplus
|
||||
|
||||
@@ -200,6 +200,7 @@ add_library(ggml-base
|
||||
ggml.cpp
|
||||
ggml-alloc.c
|
||||
ggml-backend.cpp
|
||||
ggml-backend-meta.cpp
|
||||
ggml-opt.cpp
|
||||
ggml-threading.cpp
|
||||
ggml-threading.h
|
||||
|
||||
@@ -1,5 +1,6 @@
|
||||
#include "ggml-alloc.h"
|
||||
#include "ggml-backend-impl.h"
|
||||
#include "ggml-backend.h"
|
||||
#include "ggml.h"
|
||||
#include "ggml-impl.h"
|
||||
#include <assert.h>
|
||||
@@ -1240,6 +1241,9 @@ size_t ggml_backend_alloc_ctx_tensors_from_buft_size(struct ggml_context * ctx,
|
||||
}
|
||||
|
||||
ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
|
||||
if (ggml_backend_buft_is_meta(buft)) {
|
||||
return ggml_backend_meta_alloc_ctx_tensors_from_buft(ctx, buft);
|
||||
}
|
||||
size_t nbytes_total = 0;
|
||||
return ggml_backend_alloc_ctx_tensors_from_buft_impl(ctx, buft, &nbytes_total, /*no_alloc =*/ false);
|
||||
}
|
||||
|
||||
@@ -2,7 +2,9 @@
|
||||
|
||||
// ggml-backend internal header
|
||||
|
||||
#include "ggml-alloc.h"
|
||||
#include "ggml-backend.h"
|
||||
#include "ggml.h"
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
@@ -90,9 +92,16 @@ extern "C" {
|
||||
void (*free)(ggml_backend_t backend);
|
||||
|
||||
// (optional) asynchronous tensor data access
|
||||
void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
|
||||
void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
|
||||
void (*set_tensor_async) (ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
|
||||
void (*get_tensor_async) (ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
|
||||
void (*set_tensor_2d_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
|
||||
void (*get_tensor_2d_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
|
||||
bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
|
||||
bool (*shfl_tensor_async)(ggml_backend_t backend_1, ggml_backend_t backend_2,
|
||||
const struct ggml_tensor * src1, const struct ggml_tensor * src2, struct ggml_tensor * dst1, struct ggml_tensor * dst2);
|
||||
|
||||
// (optional) backend-specific AllReduce operation for meta backend
|
||||
bool (*allreduce_tensor_async)(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends);
|
||||
|
||||
// (optional) complete all pending operations (required if the backend supports async operations)
|
||||
void (*synchronize)(ggml_backend_t backend);
|
||||
@@ -250,6 +259,9 @@ extern "C" {
|
||||
# define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
|
||||
#endif
|
||||
|
||||
// temporary workaround to statically allocate tensors from a context in a deduplicated way:
|
||||
GGML_API struct ggml_backend_buffer * ggml_backend_meta_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
+31
-11
@@ -123,7 +123,7 @@ size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
|
||||
void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
|
||||
GGML_ASSERT(buffer);
|
||||
// get_base is optional if the buffer is zero-sized
|
||||
if (buffer->size == 0) {
|
||||
if (!ggml_backend_buffer_is_meta(buffer) && buffer->size == 0) {
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@@ -388,7 +388,7 @@ ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
|
||||
|
||||
// backend copy
|
||||
|
||||
void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
|
||||
void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor * dst) {
|
||||
GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
|
||||
|
||||
if (src == dst) {
|
||||
@@ -402,7 +402,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
|
||||
} else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
|
||||
#ifndef NDEBUG
|
||||
GGML_LOG_DEBUG("%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
|
||||
#endif
|
||||
#endif // NDEBUG
|
||||
size_t nbytes = ggml_nbytes(src);
|
||||
void * data = malloc(nbytes);
|
||||
ggml_backend_tensor_get(src, data, 0, nbytes);
|
||||
@@ -411,7 +411,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
|
||||
}
|
||||
}
|
||||
|
||||
void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
|
||||
void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst) {
|
||||
GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
|
||||
|
||||
if (src == dst) {
|
||||
@@ -432,6 +432,20 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
|
||||
ggml_backend_tensor_copy(src, dst);
|
||||
}
|
||||
|
||||
void ggml_backend_tensor_shfl_async(
|
||||
ggml_backend_t backend_1, ggml_backend_t backend_2,
|
||||
const struct ggml_tensor * src1, const struct ggml_tensor * src2,
|
||||
struct ggml_tensor * dst1, struct ggml_tensor * dst2) {
|
||||
GGML_ASSERT(ggml_are_same_layout(src1, dst1) && "cannot shuffle tensors with different layouts");
|
||||
GGML_ASSERT(ggml_are_same_layout(src2, dst2) && "cannot shuffle tensors with different layouts");
|
||||
if (backend_1->iface.shfl_tensor_async != NULL) {
|
||||
if (backend_1->iface.shfl_tensor_async(backend_1, backend_2, src1, src2, dst1, dst2)) {
|
||||
return;
|
||||
}
|
||||
}
|
||||
ggml_backend_tensor_copy_async(backend_1, backend_2, src1, dst2);
|
||||
ggml_backend_tensor_copy_async(backend_2, backend_1, src2, dst1);
|
||||
}
|
||||
// events
|
||||
|
||||
ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device) {
|
||||
@@ -500,6 +514,7 @@ enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
|
||||
}
|
||||
|
||||
void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) {
|
||||
GGML_ASSERT(device);
|
||||
memset(props, 0, sizeof(*props));
|
||||
device->iface.get_props(device, props);
|
||||
}
|
||||
@@ -1455,6 +1470,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
|
||||
int split_backend_id = split->backend_id;
|
||||
ggml_backend_t split_backend = sched->backends[split_backend_id];
|
||||
|
||||
if (sched->events[split_backend_id][sched->cur_copy] == NULL) {
|
||||
ggml_backend_synchronize(split_backend);
|
||||
}
|
||||
|
||||
// copy the input tensors to the split backend
|
||||
for (int input_id = 0; input_id < split->n_inputs; input_id++) {
|
||||
ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
|
||||
@@ -1465,16 +1484,12 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
|
||||
// inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
|
||||
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
|
||||
ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
|
||||
} else {
|
||||
ggml_backend_synchronize(split_backend);
|
||||
}
|
||||
ggml_backend_tensor_copy(input, input_cpy);
|
||||
ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
|
||||
} else {
|
||||
// wait for the split backend to finish using the input before overwriting it
|
||||
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
|
||||
ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
|
||||
} else {
|
||||
ggml_backend_synchronize(split_backend);
|
||||
}
|
||||
|
||||
// when offloading MoE weights, we can reduce the amount of data copied by copying only the experts that are used
|
||||
@@ -1578,6 +1593,10 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
|
||||
}
|
||||
}
|
||||
|
||||
if (sched->events[split_backend_id][sched->cur_copy] == NULL) {
|
||||
ggml_backend_synchronize(split_backend);
|
||||
}
|
||||
|
||||
if (!sched->callback_eval) {
|
||||
enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
|
||||
if (ec != GGML_STATUS_SUCCESS) {
|
||||
@@ -1899,8 +1918,9 @@ enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct
|
||||
GGML_ASSERT(tensor->data == NULL);
|
||||
GGML_ASSERT(tensor->view_src == NULL);
|
||||
GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
|
||||
GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
|
||||
(char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
|
||||
GGML_ASSERT(ggml_backend_buffer_is_meta(buffer) ||
|
||||
(char *) addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
|
||||
(char *) ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
|
||||
|
||||
tensor->buffer = buffer;
|
||||
tensor->data = addr;
|
||||
|
||||
@@ -260,8 +260,12 @@ static struct ggml_backend_i blas_backend_i = {
|
||||
/* .get_name = */ ggml_backend_blas_get_name,
|
||||
/* .free = */ ggml_backend_blas_free,
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ NULL,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -2582,7 +2582,11 @@ static const ggml_backend_i ggml_backend_cann_interface = {
|
||||
/* .free = */ ggml_backend_cann_free,
|
||||
/* .set_tensor_async = */ ggml_backend_cann_set_tensor_async,
|
||||
/* .get_tensor_async = */ ggml_backend_cann_get_tensor_async,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ ggml_backend_cann_cpy_tensor_async,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_cann_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -268,9 +268,9 @@ static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const
|
||||
_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
|
||||
}
|
||||
|
||||
static inline __m256 quad_mx_delta_float(const int8_t x0, const float y0, const int8_t x1, const float y1) {
|
||||
return _mm256_set_m128(_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
|
||||
_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
|
||||
static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const uint8_t x1, const float y1) {
|
||||
return _mm256_set_m128(_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
|
||||
_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
|
||||
}
|
||||
#endif
|
||||
#elif defined(__SSSE3__)
|
||||
@@ -782,6 +782,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
|
||||
__m256 accum1 = _mm256_setzero_ps();
|
||||
__m256 accum2 = _mm256_setzero_ps();
|
||||
|
||||
for (; ib + 1 < nb; ib += 2) {
|
||||
const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
|
||||
const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
|
||||
@@ -795,10 +796,10 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
|
||||
const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
|
||||
const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
|
||||
accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 0].e)),
|
||||
_mm256_cvtepi32_ps(p_1), accum1);
|
||||
accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 1].e)),
|
||||
_mm256_cvtepi32_ps(p_2), accum2);
|
||||
const __m256 scale0 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 0].e));
|
||||
const __m256 scale1 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 1].e));
|
||||
accum1 = _mm256_fmadd_ps(scale0, _mm256_cvtepi32_ps(p_1), accum1);
|
||||
accum2 = _mm256_fmadd_ps(scale1, _mm256_cvtepi32_ps(p_2), accum2);
|
||||
}
|
||||
|
||||
sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
|
||||
@@ -830,7 +831,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
|
||||
|
||||
#endif
|
||||
for (; ib < nb; ++ib) {
|
||||
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
|
||||
const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib].e);
|
||||
int sumi1 = 0;
|
||||
int sumi2 = 0;
|
||||
for (int j = 0; j < QK_MXFP4/2; ++j) {
|
||||
@@ -3817,4 +3818,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
|
||||
ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
|
||||
#endif
|
||||
}
|
||||
|
||||
|
||||
@@ -75,6 +75,9 @@
|
||||
// precomputed f32 table for f16 (256 KB) (simd-mappings.h)
|
||||
float ggml_table_f32_f16[1 << 16];
|
||||
|
||||
// precomputed f32 table for e8m0 half (1 KB) (simd-mappings.h)
|
||||
float ggml_table_f32_e8m0_half[1 << 8];
|
||||
|
||||
#if defined(__ARM_ARCH)
|
||||
struct ggml_arm_arch_features_type {
|
||||
int sve_cnt;
|
||||
@@ -3681,6 +3684,11 @@ void ggml_cpu_init(void) {
|
||||
ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
|
||||
}
|
||||
|
||||
// initialize E8M0 half table (256 entries)
|
||||
for (int i = 0; i < (1 << 8); ++i) {
|
||||
ggml_table_f32_e8m0_half[i] = GGML_E8M0_TO_FP32_HALF(i);
|
||||
}
|
||||
|
||||
const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
|
||||
|
||||
GGML_PRINT_DEBUG("%s: GELU, Quick GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0);
|
||||
|
||||
@@ -195,7 +195,11 @@ static const struct ggml_backend_i ggml_backend_cpu_i = {
|
||||
/* .free = */ ggml_backend_cpu_free,
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ NULL,
|
||||
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
|
||||
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
|
||||
|
||||
@@ -116,6 +116,17 @@ extern "C" {
|
||||
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
|
||||
extern float ggml_table_f32_f16[1 << 16];
|
||||
|
||||
// precomputed f32 table for e8m0 half (1 KB)
|
||||
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
|
||||
extern float ggml_table_f32_e8m0_half[1 << 8];
|
||||
|
||||
// Use lookup table for E8M0 on x86 (faster than bit manipulation)
|
||||
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
||||
#define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
|
||||
#else
|
||||
#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
|
||||
#endif
|
||||
|
||||
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
|
||||
// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
|
||||
// This is also true for POWER9.
|
||||
|
||||
+104
-73
@@ -309,6 +309,19 @@ static ggml_cuda_device_info ggml_cuda_init() {
|
||||
// configure logging to stdout
|
||||
// CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
|
||||
|
||||
for (int id = 0; id < info.device_count; ++id) {
|
||||
ggml_cuda_set_device(id);
|
||||
for (int id_other = 0; id_other < info.device_count; ++id_other) {
|
||||
if (id == id_other) {
|
||||
continue;
|
||||
}
|
||||
int can_access_peer;
|
||||
CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
|
||||
if (can_access_peer) {
|
||||
CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
|
||||
}
|
||||
}
|
||||
}
|
||||
return info;
|
||||
}
|
||||
|
||||
@@ -1371,64 +1384,6 @@ static void ggml_cuda_op_mul_mat_cublas(
|
||||
GGML_UNUSED_VARS(dst, src1_ddq_i, src1_padded_row_size);
|
||||
}
|
||||
|
||||
static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
|
||||
static bool peer_access_enabled = false;
|
||||
|
||||
const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;
|
||||
|
||||
if (peer_access_enabled == enable_peer_access) {
|
||||
return;
|
||||
}
|
||||
|
||||
#ifdef NDEBUG
|
||||
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
|
||||
ggml_cuda_set_device(id);
|
||||
CUDA_CHECK(cudaDeviceSynchronize());
|
||||
}
|
||||
|
||||
for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
|
||||
ggml_cuda_set_device(id);
|
||||
|
||||
for (int id_other = 0; id_other < ggml_backend_cuda_get_device_count(); ++id_other) {
|
||||
if (id == id_other) {
|
||||
continue;
|
||||
}
|
||||
if (id != main_device && id_other != main_device) {
|
||||
continue;
|
||||
}
|
||||
|
||||
int can_access_peer;
|
||||
CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
|
||||
if (can_access_peer) {
|
||||
if (enable_peer_access) {
|
||||
cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
|
||||
if (err != cudaErrorPeerAccessAlreadyEnabled) {
|
||||
CUDA_CHECK(err);
|
||||
} else {
|
||||
// reset the error
|
||||
(void)cudaGetLastError();
|
||||
}
|
||||
} else {
|
||||
cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
|
||||
if (err != cudaErrorPeerAccessNotEnabled) {
|
||||
CUDA_CHECK(err);
|
||||
} else {
|
||||
// reset the error
|
||||
(void)cudaGetLastError();
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
ggml_cuda_set_device(main_device);
|
||||
#endif // NDEBUG
|
||||
|
||||
peer_access_enabled = enable_peer_access;
|
||||
|
||||
GGML_UNUSED(main_device);
|
||||
}
|
||||
|
||||
static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
|
||||
void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
|
||||
|
||||
@@ -2420,11 +2375,6 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
|
||||
}
|
||||
|
||||
static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
|
||||
// why is this here instead of mul_mat?
|
||||
if (dst->src[0] != nullptr && ggml_backend_buft_is_cuda_split(dst->src[0]->buffer->buft)) {
|
||||
ggml_cuda_set_peer_access(dst->src[1]->ne[1], ctx.device);
|
||||
}
|
||||
|
||||
switch (dst->op) {
|
||||
case GGML_OP_ARGMAX:
|
||||
ggml_cuda_argmax(ctx, dst);
|
||||
@@ -2800,29 +2750,35 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
|
||||
ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
|
||||
ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
|
||||
|
||||
if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
|
||||
//enables async copies from CPU to CUDA, instead of only CUDA-to-CUDA
|
||||
bool copy_from_host = ggml_backend_buffer_is_host(buf_src) && ggml_backend_dev_type(backend_src->device) == GGML_BACKEND_DEVICE_TYPE_CPU;
|
||||
|
||||
if (!(copy_from_host || ggml_backend_is_cuda(backend_src)) || !ggml_backend_is_cuda(backend_dst)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
|
||||
if (!(copy_from_host || ggml_backend_buffer_is_cuda(buf_src)) || !ggml_backend_buffer_is_cuda(buf_dst)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// device -> device copy
|
||||
ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
|
||||
ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
|
||||
ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *) backend_src->context;
|
||||
ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *) backend_dst->context;
|
||||
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *) buf_src->context;
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *) buf_dst->context;
|
||||
|
||||
if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
|
||||
if ((copy_from_host && cuda_ctx_dst->device != buf_ctx_dst->device) ||
|
||||
!copy_from_host && (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device)) {
|
||||
#ifndef NDEBUG
|
||||
GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
|
||||
#endif
|
||||
#endif // NDEBUG
|
||||
return false;
|
||||
}
|
||||
|
||||
if (backend_src != backend_dst) {
|
||||
if (copy_from_host) {
|
||||
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyHostToDevice, cuda_ctx_dst->stream()));
|
||||
} else if (backend_src != backend_dst) {
|
||||
// copy on src stream
|
||||
if (cuda_ctx_src->device == cuda_ctx_dst->device) {
|
||||
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
|
||||
@@ -2831,7 +2787,7 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
|
||||
return false;
|
||||
#else
|
||||
CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), cuda_ctx_src->stream()));
|
||||
#endif
|
||||
#endif // GGML_CUDA_NO_PEER_COPY
|
||||
}
|
||||
|
||||
// record event on src stream after the copy
|
||||
@@ -2851,6 +2807,77 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
|
||||
return true;
|
||||
}
|
||||
|
||||
static bool ggml_backend_cuda_shfl_tensor_async(
|
||||
ggml_backend_t backend_1, ggml_backend_t backend_2, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst1, ggml_tensor * dst2) {
|
||||
ggml_backend_buffer_t buf_src1 = src1->view_src ? src1->view_src->buffer : src1->buffer;
|
||||
ggml_backend_buffer_t buf_src2 = src2->view_src ? src2->view_src->buffer : src2->buffer;
|
||||
ggml_backend_buffer_t buf_dst1 = dst1->view_src ? dst1->view_src->buffer : dst1->buffer;
|
||||
ggml_backend_buffer_t buf_dst2 = dst2->view_src ? dst2->view_src->buffer : dst2->buffer;
|
||||
|
||||
if (!ggml_backend_is_cuda(backend_1) || !ggml_backend_is_cuda(backend_2)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!ggml_backend_buffer_is_cuda(buf_src1) || !ggml_backend_buffer_is_cuda(buf_src2) ||
|
||||
!ggml_backend_buffer_is_cuda(buf_dst1) || !ggml_backend_buffer_is_cuda(buf_dst2)) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// device -> device copy
|
||||
ggml_backend_cuda_context * cuda_ctx_1 = (ggml_backend_cuda_context *) backend_1->context;
|
||||
ggml_backend_cuda_context * cuda_ctx_2 = (ggml_backend_cuda_context *) backend_2->context;
|
||||
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_src1 = (ggml_backend_cuda_buffer_context *) buf_src1->context;
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_src2 = (ggml_backend_cuda_buffer_context *) buf_src2->context;
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_dst1 = (ggml_backend_cuda_buffer_context *) buf_dst1->context;
|
||||
ggml_backend_cuda_buffer_context * buf_ctx_dst2 = (ggml_backend_cuda_buffer_context *) buf_dst2->context;
|
||||
|
||||
if (cuda_ctx_1->device != buf_ctx_src1->device || cuda_ctx_2->device != buf_ctx_src2->device ||
|
||||
cuda_ctx_1->device != buf_ctx_dst1->device || cuda_ctx_2->device != buf_ctx_dst2->device) {
|
||||
#ifndef NDEBUG
|
||||
GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
|
||||
#endif // NDEBUG
|
||||
return false;
|
||||
}
|
||||
|
||||
if (backend_1 != backend_2) {
|
||||
// Copies under control of src streams:
|
||||
if (cuda_ctx_1->device == cuda_ctx_2->device) {
|
||||
CUDA_CHECK(cudaMemcpyAsync(dst2->data, src1->data, ggml_nbytes(dst2), cudaMemcpyDeviceToDevice, cuda_ctx_1->stream()));
|
||||
CUDA_CHECK(cudaMemcpyAsync(dst1->data, src2->data, ggml_nbytes(dst1), cudaMemcpyDeviceToDevice, cuda_ctx_2->stream()));
|
||||
} else {
|
||||
#ifdef GGML_CUDA_NO_PEER_COPY
|
||||
return false;
|
||||
#else
|
||||
CUDA_CHECK(cudaMemcpyPeerAsync(dst2->data, cuda_ctx_2->device, src1->data, cuda_ctx_1->device, ggml_nbytes(dst2), cuda_ctx_1->stream()));
|
||||
CUDA_CHECK(cudaMemcpyPeerAsync(dst1->data, cuda_ctx_1->device, src2->data, cuda_ctx_2->device, ggml_nbytes(dst1), cuda_ctx_2->stream()));
|
||||
#endif // GGML_CUDA_NO_PEER_COPY
|
||||
}
|
||||
|
||||
// Record event on src streams after the copy:
|
||||
if (!cuda_ctx_1->copy_event) {
|
||||
ggml_cuda_set_device(cuda_ctx_1->device);
|
||||
CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_1->copy_event, cudaEventDisableTiming));
|
||||
}
|
||||
if (!cuda_ctx_2->copy_event) {
|
||||
ggml_cuda_set_device(cuda_ctx_2->device);
|
||||
CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_2->copy_event, cudaEventDisableTiming));
|
||||
}
|
||||
|
||||
CUDA_CHECK(cudaEventRecord(cuda_ctx_1->copy_event, cuda_ctx_1->stream()));
|
||||
CUDA_CHECK(cudaEventRecord(cuda_ctx_2->copy_event, cuda_ctx_2->stream()));
|
||||
|
||||
// Wait on dst stream for the copies to complete:
|
||||
CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_2->stream(), cuda_ctx_1->copy_event, 0));
|
||||
CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_1->stream(), cuda_ctx_2->copy_event, 0));
|
||||
} else {
|
||||
// srcs and dsts are on the same backend:
|
||||
CUDA_CHECK(cudaMemcpyAsync(dst2->data, src1->data, ggml_nbytes(dst2), cudaMemcpyDeviceToDevice, cuda_ctx_1->stream()));
|
||||
CUDA_CHECK(cudaMemcpyAsync(dst1->data, src2->data, ggml_nbytes(dst1), cudaMemcpyDeviceToDevice, cuda_ctx_2->stream()));
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
|
||||
ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
|
||||
|
||||
@@ -4251,7 +4278,11 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
|
||||
/* .free = */ ggml_backend_cuda_free,
|
||||
/* .set_tensor_async = */ ggml_backend_cuda_set_tensor_async,
|
||||
/* .get_tensor_async = */ ggml_backend_cuda_get_tensor_async,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ ggml_backend_cuda_cpy_tensor_async,
|
||||
/* .shfl_tensor_async = */ ggml_backend_cuda_shfl_tensor_async,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_cuda_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -2756,7 +2756,11 @@ static struct ggml_backend_i hexagon_backend_i = {
|
||||
/* .free = */ ggml_backend_hexagon_free,
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_hexagon_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -394,7 +394,7 @@ bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, con
|
||||
[encoder endEncoding];
|
||||
|
||||
ggml_metal_event_t ev_cpy = ggml_metal_get_ev_cpy(ctx_src);
|
||||
ggml_metal_event_record(ctx_src, ev_cpy);
|
||||
ggml_metal_event_encode_signal(ev_cpy, cmd_buf);
|
||||
|
||||
[cmd_buf commit];
|
||||
|
||||
@@ -415,7 +415,7 @@ bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, con
|
||||
|
||||
enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph * gf) {
|
||||
// number of nodes encoded by the main thread (empirically determined)
|
||||
const int n_main = 64;
|
||||
const int n_main = MAX(64, 0.1*gf->n_nodes);
|
||||
|
||||
// number of threads in addition to the main thread
|
||||
const int n_cb = ctx->n_cb;
|
||||
|
||||
@@ -176,6 +176,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows(ggml_me
|
||||
return res;
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag(ggml_metal_library_t lib, const ggml_tensor * op) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
|
||||
const int n = op->src[0]->ne[0];
|
||||
|
||||
snprintf(base, 256, "kernel_diag_%s", ggml_type_name(op->src[0]->type));
|
||||
snprintf(name, 256, "%s_n=%d", base, n);
|
||||
|
||||
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
|
||||
if (!res.pipeline) {
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
|
||||
}
|
||||
|
||||
res.nsg = 1;
|
||||
res.smem = 0;
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat(ggml_metal_library_t lib, ggml_type tsrc) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
@@ -534,6 +554,36 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv(ggml_metal_
|
||||
return res;
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
|
||||
const int nsg = 8;
|
||||
const int n = op->src[1]->ne[1];
|
||||
const int k = op->src[1]->ne[0];
|
||||
|
||||
snprintf(base, 256, "kernel_solve_tri_%s", ggml_type_name(op->src[0]->type));
|
||||
snprintf(name, 256, "%s_nsg=%d_n=%d_k=%d", base, nsg, n, k);
|
||||
|
||||
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
|
||||
if (!res.pipeline) {
|
||||
ggml_metal_cv_t cv = ggml_metal_cv_init();
|
||||
|
||||
ggml_metal_cv_set_int16(cv, nsg, FC_SOLVE_TRI + 0);
|
||||
ggml_metal_cv_set_int16(cv, n, FC_SOLVE_TRI + 1);
|
||||
ggml_metal_cv_set_int16(cv, k, FC_SOLVE_TRI + 2);
|
||||
|
||||
res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
|
||||
|
||||
ggml_metal_cv_free(cv);
|
||||
}
|
||||
|
||||
res.nsg = nsg;
|
||||
res.smem = GGML_PAD(GGML_PAD(n, 32)*nsg*sizeof(float), 16);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int nsg, int nxpsg, int r1ptg) {
|
||||
char base[256];
|
||||
char name[256];
|
||||
|
||||
@@ -108,6 +108,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_1d
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_2d (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_get_rows (ggml_metal_library_t lib, enum ggml_type tsrc);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat (ggml_metal_library_t lib, enum ggml_type tsrc);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
@@ -121,6 +122,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv (ggml_metal_library_t lib, const struct ggml_tensor * op);
|
||||
|
||||
@@ -1153,6 +1153,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
case GGML_OP_RWKV_WKV6:
|
||||
case GGML_OP_RWKV_WKV7:
|
||||
return true;
|
||||
case GGML_OP_SOLVE_TRI:
|
||||
case GGML_OP_MUL_MAT:
|
||||
case GGML_OP_MUL_MAT_ID:
|
||||
return has_simdgroup_reduction;
|
||||
@@ -1234,6 +1235,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
|
||||
return false;
|
||||
};
|
||||
}
|
||||
case GGML_OP_DIAG:
|
||||
return true;
|
||||
case GGML_OP_OPT_STEP_ADAMW:
|
||||
case GGML_OP_OPT_STEP_SGD:
|
||||
return has_simdgroup_reduction;
|
||||
|
||||
@@ -78,7 +78,8 @@
|
||||
#define FC_MUL_MM 700
|
||||
#define FC_ROPE 800
|
||||
#define FC_SSM_CONV 900
|
||||
#define FC_COUNT_EQUAL 1000
|
||||
#define FC_SOLVE_TRI 1000
|
||||
#define FC_COUNT_EQUAL 1100
|
||||
|
||||
// op-specific constants
|
||||
#define OP_FLASH_ATTN_EXT_NQPSG 8
|
||||
@@ -733,6 +734,33 @@ typedef struct {
|
||||
uint64_t nb0;
|
||||
} ggml_metal_kargs_ssm_scan;
|
||||
|
||||
typedef struct {
|
||||
int32_t ne00;
|
||||
int32_t ne01;
|
||||
int32_t ne02;
|
||||
int32_t ne03;
|
||||
uint64_t nb00;
|
||||
uint64_t nb01;
|
||||
uint64_t nb02;
|
||||
uint64_t nb03;
|
||||
int32_t ne10;
|
||||
int32_t ne11;
|
||||
int32_t ne12;
|
||||
int32_t ne13;
|
||||
uint64_t nb10;
|
||||
uint64_t nb11;
|
||||
uint64_t nb12;
|
||||
uint64_t nb13;
|
||||
int32_t ne0;
|
||||
int32_t ne1;
|
||||
int32_t ne2;
|
||||
int32_t ne3;
|
||||
uint64_t nb0;
|
||||
uint64_t nb1;
|
||||
uint64_t nb2;
|
||||
uint64_t nb3;
|
||||
} ggml_metal_kargs_solve_tri;
|
||||
|
||||
typedef struct {
|
||||
int32_t ne00t;
|
||||
int32_t ne00;
|
||||
@@ -764,6 +792,25 @@ typedef struct {
|
||||
uint64_t nb3;
|
||||
} ggml_metal_kargs_set_rows;
|
||||
|
||||
typedef struct {
|
||||
int32_t ne00;
|
||||
int32_t ne01;
|
||||
int32_t ne02;
|
||||
int32_t ne03;
|
||||
uint64_t nb00;
|
||||
uint64_t nb01;
|
||||
uint64_t nb02;
|
||||
uint64_t nb03;
|
||||
int32_t ne0;
|
||||
int32_t ne1;
|
||||
int32_t ne2;
|
||||
int32_t ne3;
|
||||
uint64_t nb0;
|
||||
uint64_t nb1;
|
||||
uint64_t nb2;
|
||||
uint64_t nb3;
|
||||
} ggml_metal_kargs_diag;
|
||||
|
||||
typedef struct {
|
||||
int64_t ne00;
|
||||
int64_t ne01;
|
||||
|
||||
@@ -341,6 +341,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
|
||||
{
|
||||
n_fuse = ggml_metal_op_rwkv(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_SOLVE_TRI:
|
||||
{
|
||||
n_fuse = ggml_metal_op_solve_tri(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_MUL_MAT:
|
||||
{
|
||||
n_fuse = ggml_metal_op_mul_mat(ctx, idx);
|
||||
@@ -357,6 +361,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
|
||||
{
|
||||
n_fuse = ggml_metal_op_set_rows(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_DIAG:
|
||||
{
|
||||
n_fuse = ggml_metal_op_diag(ctx, idx);
|
||||
} break;
|
||||
case GGML_OP_L2_NORM:
|
||||
{
|
||||
n_fuse = ggml_metal_op_l2_norm(ctx, idx);
|
||||
@@ -1255,6 +1263,48 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
ggml_metal_library_t lib = ctx->lib;
|
||||
ggml_metal_encoder_t enc = ctx->enc;
|
||||
|
||||
GGML_TENSOR_LOCALS(int32_t, ne0, op->src[0], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
|
||||
GGML_TENSOR_LOCALS(int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
ggml_metal_kargs_diag args = {
|
||||
/*.ne00 =*/ne00,
|
||||
/*.ne01 =*/ne01,
|
||||
/*.ne02 =*/ne02,
|
||||
/*.ne03 =*/ne03,
|
||||
/*.nb00 =*/nb00,
|
||||
/*.nb01 =*/nb01,
|
||||
/*.nb02 =*/nb02,
|
||||
/*.nb03 =*/nb03,
|
||||
/*.ne0 =*/ne0,
|
||||
/*.ne1 =*/ne1,
|
||||
/*.ne2 =*/ne2,
|
||||
/*.ne3 =*/ne3,
|
||||
/*.nb0 =*/nb0,
|
||||
/*.nb1 =*/nb1,
|
||||
/*.nb2 =*/nb2,
|
||||
/*.nb3 =*/nb3,
|
||||
};
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_diag(lib, op);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
|
||||
ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 2);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, 32, 1, 1);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
@@ -1557,6 +1607,63 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
ggml_metal_library_t lib = ctx->lib;
|
||||
ggml_metal_encoder_t enc = ctx->enc;
|
||||
|
||||
GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
|
||||
GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
|
||||
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
|
||||
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
|
||||
|
||||
ggml_metal_kargs_solve_tri args = {
|
||||
/*.ne00 =*/ ne00,
|
||||
/*.ne01 =*/ ne01,
|
||||
/*.ne02 =*/ ne02,
|
||||
/*.ne03 =*/ ne03,
|
||||
/*.nb00 =*/ nb00,
|
||||
/*.nb01 =*/ nb01,
|
||||
/*.nb02 =*/ nb02,
|
||||
/*.nb03 =*/ nb03,
|
||||
/*.ne10 =*/ ne10,
|
||||
/*.ne11 =*/ ne11,
|
||||
/*.ne12 =*/ ne12,
|
||||
/*.ne13 =*/ ne13,
|
||||
/*.nb10 =*/ nb10,
|
||||
/*.nb11 =*/ nb11,
|
||||
/*.nb12 =*/ nb12,
|
||||
/*.nb13 =*/ nb13,
|
||||
/*.ne0 =*/ ne0,
|
||||
/*.ne1 =*/ ne1,
|
||||
/*.ne2 =*/ ne2,
|
||||
/*.ne3 =*/ ne3,
|
||||
/*.nb0 =*/ nb0,
|
||||
/*.nb1 =*/ nb1,
|
||||
/*.nb2 =*/ nb2,
|
||||
/*.nb3 =*/ nb3,
|
||||
};
|
||||
|
||||
auto pipeline = ggml_metal_library_get_pipeline_solve_tri(lib, op);
|
||||
|
||||
ggml_metal_encoder_set_pipeline(enc, pipeline);
|
||||
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
|
||||
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 3);
|
||||
|
||||
const int nsg = pipeline.nsg;
|
||||
|
||||
ggml_metal_encoder_set_threadgroup_memory_size(enc, pipeline.smem, 0);
|
||||
|
||||
ggml_metal_encoder_dispatch_threadgroups(enc, (ne10 + nsg - 1)/nsg, ne02, ne03, 32, nsg, 1);
|
||||
|
||||
return 1;
|
||||
}
|
||||
|
||||
int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
|
||||
ggml_tensor * op = ctx->node(idx);
|
||||
|
||||
|
||||
@@ -56,10 +56,12 @@ int ggml_metal_op_sum_rows (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_cumsum (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_get_rows (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_set_rows (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_diag (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_soft_max (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_ssm_conv (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_ssm_scan (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_rwkv (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_solve_tri (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_cpy (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_pool_1d (ggml_metal_op_t ctx, int idx);
|
||||
int ggml_metal_op_pool_2d (ggml_metal_op_t ctx, int idx);
|
||||
|
||||
@@ -7,6 +7,9 @@
|
||||
#include "ggml-metal-context.h"
|
||||
#include "ggml-metal-ops.h"
|
||||
|
||||
#include <mutex>
|
||||
#include <string>
|
||||
|
||||
#define GGML_METAL_NAME "MTL"
|
||||
#define GGML_METAL_MAX_DEVICES 16
|
||||
|
||||
@@ -560,7 +563,11 @@ static ggml_backend_i ggml_backend_metal_i = {
|
||||
/* .free = */ ggml_backend_metal_free,
|
||||
/* .set_tensor_async = */ ggml_backend_metal_set_tensor_async,
|
||||
/* .get_tensor_async = */ ggml_backend_metal_get_tensor_async,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ ggml_backend_metal_cpy_tensor_async, // only needed for multi-GPU setups
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_metal_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -2737,6 +2737,83 @@ kernel void kernel_rwkv_wkv7_f32(
|
||||
}
|
||||
}
|
||||
|
||||
constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
|
||||
constant short FC_solve_tri_n [[function_constant(FC_SOLVE_TRI + 1)]];
|
||||
constant short FC_solve_tri_k [[function_constant(FC_SOLVE_TRI + 2)]];
|
||||
|
||||
kernel void kernel_solve_tri_f32(
|
||||
constant ggml_metal_kargs_solve_tri & args,
|
||||
device const char * src0,
|
||||
device const char * src1,
|
||||
device char * dst,
|
||||
threadgroup char * shmem [[threadgroup(0)]],
|
||||
ushort3 tgpig[[threadgroup_position_in_grid]],
|
||||
ushort sgitg[[simdgroup_index_in_threadgroup]],
|
||||
ushort tiisg[[thread_index_in_simdgroup]],
|
||||
ushort3 ntg[[threads_per_threadgroup]]) {
|
||||
constexpr short NW = N_SIMDWIDTH;
|
||||
|
||||
const short NSG = FC_solve_tri_nsg;
|
||||
const short N = FC_solve_tri_n;
|
||||
const short K = FC_solve_tri_k;
|
||||
const short NP = PAD2(N, NW);
|
||||
|
||||
const int32_t ne02 = args.ne02;
|
||||
const int32_t ne03 = args.ne03;
|
||||
|
||||
const int32_t i03 = tgpig.z;
|
||||
const int32_t i02 = tgpig.y;
|
||||
const int32_t i01 = tgpig.x*NSG + sgitg;
|
||||
|
||||
threadgroup float * sh0 = (threadgroup float *) shmem;
|
||||
|
||||
device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
|
||||
device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
|
||||
device float * dst_ptr = (device float *)(dst + i02 * args.nb2 + i03 * args.nb3) + i01;
|
||||
|
||||
for (short rr = 0; rr < N; rr += NSG) {
|
||||
threadgroup_barrier(mem_flags::mem_threadgroup);
|
||||
|
||||
{
|
||||
threadgroup float * sh0_cur = sh0 + sgitg*NP;
|
||||
|
||||
for (short t = 0; t*NW < N; ++t) {
|
||||
const short idx = t*NW + tiisg;
|
||||
sh0_cur[idx] = src0_ptr[idx];
|
||||
}
|
||||
|
||||
src0_ptr += NSG*N;
|
||||
}
|
||||
|
||||
threadgroup_barrier(mem_flags::mem_threadgroup);
|
||||
|
||||
if (i01 >= args.ne10) {
|
||||
continue;
|
||||
}
|
||||
|
||||
for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
|
||||
const short r = rr + ir;
|
||||
|
||||
threadgroup float * sh0_cur = sh0 + ir*NP;
|
||||
|
||||
float sum = 0.0f;
|
||||
|
||||
for (short t = 0; t*NW < r; ++t) {
|
||||
const short idx = t*NW + tiisg;
|
||||
sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
|
||||
}
|
||||
|
||||
sum = simd_sum(sum);
|
||||
|
||||
if (tiisg == 0) {
|
||||
const float diag = sh0_cur[r];
|
||||
|
||||
dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_argmax_f32(
|
||||
constant ggml_metal_kargs_argmax & args,
|
||||
device const char * src0,
|
||||
@@ -8738,6 +8815,26 @@ kernel void kernel_set_rows_f(
|
||||
}
|
||||
}
|
||||
|
||||
kernel void kernel_diag_f32(
|
||||
constant ggml_metal_kargs_diag & args,
|
||||
device const char * src0,
|
||||
device char * dst,
|
||||
uint3 tgpig[[threadgroup_position_in_grid]],
|
||||
ushort tiitg[[thread_index_in_threadgroup]]) {
|
||||
constexpr short NW = N_SIMDWIDTH;
|
||||
|
||||
const int32_t i3 = tgpig.z;
|
||||
const int32_t i2 = tgpig.y;
|
||||
const int32_t i1 = tgpig.x;
|
||||
|
||||
device const float * src0_ptr = (device const float *)(src0 + i2*args.nb02 + i3*args.nb03);
|
||||
device float * dst_ptr = (device float *)(dst + i1*args.nb01 + i2*args.nb2 + i3*args.nb3);
|
||||
|
||||
for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
|
||||
dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
|
||||
}
|
||||
}
|
||||
|
||||
constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
|
||||
constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
|
||||
|
||||
|
||||
@@ -3478,6 +3478,10 @@ static ggml_backend_i ggml_backend_opencl_i = {
|
||||
/* .set_tensor_async = */ NULL, /* ggml_backend_opencl_set_tensor_async */
|
||||
/* .get_tensor_async = */ NULL, /* ggml_backend_opencl_get_tensor_async */
|
||||
/* .cpy_tensor_async = */ NULL, /* ggml_backend_opencl_cpy_tensor_async */
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_opencl_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -893,6 +893,10 @@ static ggml_backend_i ggml_backend_rpc_interface = {
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_rpc_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -4455,9 +4455,13 @@ static ggml_backend_i ggml_backend_sycl_interface = {
|
||||
/* .free = */ ggml_backend_sycl_free,
|
||||
/* .set_tensor_async = */ ggml_backend_sycl_set_tensor_async,
|
||||
/* .get_tensor_async = */ ggml_backend_sycl_get_tensor_async,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL, // ggml_backend_sycl_cpy_tensor_async,
|
||||
// // TODO: update for the new
|
||||
// interface
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_sycl_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -36,7 +36,7 @@ apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor) {
|
||||
result.data = reinterpret_cast<uint64_t>(tensor->data);
|
||||
if (tensor->data) {
|
||||
if (!tensor->buffer) {
|
||||
GGML_ABORT("tensor has data but not buffer");
|
||||
GGML_ABORT("%s: tensor has data but not buffer", __func__);
|
||||
}
|
||||
// tensor->data is serialized as an offset to the buffer base address
|
||||
result.data -= reinterpret_cast<uint64_t>(BUFFER_TO_GGML_CONTEXT(tensor->buffer)->base);
|
||||
|
||||
@@ -27,7 +27,7 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
|
||||
|
||||
const void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
if (!shmem_data) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
return 1;
|
||||
}
|
||||
@@ -45,7 +45,7 @@ uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, v
|
||||
if (dev->iface.supports_op(dev, op)) {
|
||||
continue;
|
||||
}
|
||||
GGML_LOG_ERROR("Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
|
||||
|
||||
status = GGML_STATUS_ABORTED;
|
||||
apir_encode_ggml_status(enc, &status);
|
||||
|
||||
@@ -36,18 +36,22 @@ uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec
|
||||
ggml_backend_buffer_type_t buft;
|
||||
buft = apir_decode_ggml_buffer_type(dec);
|
||||
|
||||
size_t value = buft->iface.get_max_size(buft);
|
||||
size_t value = SIZE_MAX;
|
||||
if (buft->iface.get_max_size) {
|
||||
value = buft->iface.get_max_size(buft);
|
||||
}
|
||||
|
||||
apir_encode_size_t(enc, &value);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
|
||||
uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
|
||||
GGML_UNUSED(ctx);
|
||||
ggml_backend_buffer_type_t buft;
|
||||
buft = apir_decode_ggml_buffer_type(dec);
|
||||
GGML_UNUSED(dec);
|
||||
const bool is_host = false;
|
||||
|
||||
bool is_host = buft->iface.is_host(buft);
|
||||
apir_encode_bool_t(enc, &is_host);
|
||||
|
||||
return 0;
|
||||
|
||||
@@ -40,7 +40,7 @@ uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl
|
||||
void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
|
||||
if (!shmem_data) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
@@ -71,7 +71,7 @@ uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl
|
||||
|
||||
void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
if (!shmem_data) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
@@ -121,7 +121,7 @@ uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virg
|
||||
buffer = apir_decode_ggml_buffer(dec);
|
||||
|
||||
if (!apir_untrack_backend_buffer(buffer)) {
|
||||
GGML_LOG_WARN("%s: unknown buffer %p\n", __func__, (void *) buffer);
|
||||
GGML_LOG_WARN(GGML_VIRTGPU_BCK "%s: unknown buffer %p\n", __func__, (void *) buffer);
|
||||
return 1;
|
||||
}
|
||||
|
||||
|
||||
@@ -124,7 +124,7 @@ uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec,
|
||||
|
||||
void * shmem_ptr = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
|
||||
if (!shmem_ptr) {
|
||||
GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: Couldn't get the shmem addr from virgl\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
return 1;
|
||||
}
|
||||
|
||||
@@ -17,26 +17,26 @@ uint64_t timer_count = 0;
|
||||
|
||||
uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p) {
|
||||
if (reg != NULL) {
|
||||
GGML_LOG_WARN("%s: already initialized\n", __func__);
|
||||
GGML_LOG_WARN(GGML_VIRTGPU_BCK "%s: already initialized\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_ALREADY_INITED;
|
||||
}
|
||||
ggml_backend_reg_t (*ggml_backend_reg_fct)(void) = (ggml_backend_reg_t (*)()) ggml_backend_reg_fct_p;
|
||||
|
||||
reg = ggml_backend_reg_fct();
|
||||
if (reg == NULL) {
|
||||
GGML_LOG_ERROR("%s: backend registration failed\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend registration failed\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED;
|
||||
}
|
||||
|
||||
if (!reg->iface.get_device_count(reg)) {
|
||||
GGML_LOG_ERROR("%s: backend initialization failed: no device found\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device found\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_NO_DEVICE;
|
||||
}
|
||||
|
||||
dev = reg->iface.get_device(reg, 0);
|
||||
|
||||
if (!dev) {
|
||||
GGML_LOG_ERROR("%s: backend initialization failed: no device received\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK "%s: backend initialization failed: no device received\n", __func__);
|
||||
return APIR_BACKEND_INITIALIZE_NO_DEVICE;
|
||||
}
|
||||
|
||||
|
||||
@@ -16,6 +16,7 @@ uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec,
|
||||
uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
|
||||
uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
|
||||
@@ -62,7 +63,7 @@ static inline const char * backend_dispatch_command_name(ApirBackendCommandType
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
|
||||
return "backend_buffer_type_get_max_size";
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
|
||||
return "backend_buffer_type_is_host";
|
||||
return "backend_buffer_type_is_host (DEPRECATED)";
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
|
||||
return "backend_buffer_type_alloc_buffer";
|
||||
case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
|
||||
@@ -110,7 +111,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME = */ backend_buffer_type_get_name,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT = */ backend_buffer_type_get_alignment,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE = */ backend_buffer_type_get_max_size,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = */ backend_buffer_type_is_host,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = */ backend_buffer_type_is_host /* DEPRECATED */,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER = */ backend_buffer_type_alloc_buffer,
|
||||
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE = */ backend_buffer_type_get_alloc_size,
|
||||
|
||||
|
||||
@@ -11,6 +11,8 @@
|
||||
#include "shared/apir_cs.h"
|
||||
#include "shared/apir_cs_ggml.h"
|
||||
|
||||
#define GGML_VIRTGPU_BCK "ggml-virtgpu-backend: "
|
||||
|
||||
struct virgl_apir_context {
|
||||
uint32_t ctx_id;
|
||||
virgl_apir_callbacks * iface;
|
||||
|
||||
@@ -35,14 +35,8 @@ void apir_backend_deinit(uint32_t virgl_ctx_id) {
|
||||
buffer->iface.free_buffer(buffer);
|
||||
}
|
||||
|
||||
if (dev) {
|
||||
size_t free, total;
|
||||
dev->iface.get_memory(dev, &free, &total);
|
||||
GGML_LOG_INFO("%s: free memory: %ld MB\n", __func__, (size_t) free / 1024 / 1024);
|
||||
}
|
||||
|
||||
if (backend_library_handle) {
|
||||
GGML_LOG_INFO("%s: The GGML backend library was loaded. Unloading it.\n", __func__);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU_BCK "The GGML backend library was loaded. Unloading it.\n");
|
||||
dlclose(backend_library_handle);
|
||||
backend_library_handle = NULL;
|
||||
}
|
||||
@@ -65,7 +59,7 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
if (apir_logfile) {
|
||||
ggml_log_set(log_to_file_callback, apir_logfile);
|
||||
} else {
|
||||
GGML_LOG_INFO("Could not open the log file at '%s'\n", apir_log_to_file);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU_BCK "Could not open the log file at '%s'\n", apir_log_to_file);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -74,7 +68,10 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
const char * library_reg = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG;
|
||||
|
||||
if (!library_name) {
|
||||
GGML_LOG_ERROR("cannot open the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot open the GGML library: env var '%s' not defined\n",
|
||||
__func__, APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
|
||||
|
||||
|
||||
return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
|
||||
}
|
||||
@@ -82,13 +79,16 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
backend_library_handle = dlopen(library_name, RTLD_LAZY);
|
||||
|
||||
if (!backend_library_handle) {
|
||||
GGML_LOG_ERROR("cannot open the GGML library: %s\n", dlerror());
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot open the GGML library: %s\n", __func__, dlerror());
|
||||
|
||||
return APIR_LOAD_LIBRARY_CANNOT_OPEN;
|
||||
}
|
||||
|
||||
if (!library_reg) {
|
||||
GGML_LOG_ERROR("cannot register the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot register the GGML library: env var '%s' not defined\n",
|
||||
__func__, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
|
||||
|
||||
return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
|
||||
}
|
||||
@@ -96,8 +96,10 @@ ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct
|
||||
void * ggml_backend_reg_fct = dlsym(backend_library_handle, library_reg);
|
||||
dlsym_error = dlerror();
|
||||
if (dlsym_error) {
|
||||
GGML_LOG_ERROR("cannot find the GGML backend registration symbol '%s' (from %s): %s\n", library_reg,
|
||||
APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: cannot find the GGML backend registration symbol '%s' (from %s): %s\n",
|
||||
__func__, library_reg, APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
|
||||
|
||||
|
||||
return APIR_LOAD_LIBRARY_SYMBOL_MISSING;
|
||||
}
|
||||
@@ -134,7 +136,9 @@ uint32_t apir_backend_dispatcher(uint32_t virgl_ctx_id,
|
||||
};
|
||||
|
||||
if (cmd_type >= APIR_BACKEND_DISPATCH_TABLE_COUNT) {
|
||||
GGML_LOG_ERROR("Received an invalid dispatch index (%d >= %d)\n", cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU_BCK
|
||||
"%s: Received an invalid dispatch index (%d >= %d)\n",
|
||||
__func__, cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
|
||||
return APIR_BACKEND_FORWARD_INDEX_INVALID;
|
||||
}
|
||||
|
||||
|
||||
@@ -86,7 +86,7 @@ static inline bool apir_decoder_peek_internal(apir_decoder * dec,
|
||||
assert(val_size <= size);
|
||||
|
||||
if (unlikely(size > (size_t) (dec->end - dec->cur))) {
|
||||
GGML_LOG_ERROR("reading too much from the decoder ...\n");
|
||||
GGML_LOG_ERROR("%s: reading too much from the decoder ...\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
memset(val, 0, val_size);
|
||||
return false;
|
||||
@@ -103,7 +103,7 @@ static inline void apir_decoder_peek(apir_decoder * dec, size_t size, void * val
|
||||
|
||||
static inline const void * apir_decoder_use_inplace(apir_decoder * dec, size_t size) {
|
||||
if (unlikely(size > (size_t) (dec->end - dec->cur))) {
|
||||
GGML_LOG_ERROR("reading too much from the decoder ...\n");
|
||||
GGML_LOG_ERROR("%s: reading too much from the decoder ...\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
return NULL;
|
||||
}
|
||||
@@ -221,7 +221,7 @@ static inline uint64_t apir_decode_array_size(apir_decoder * dec, uint64_t expec
|
||||
uint64_t size;
|
||||
apir_decode_uint64_t(dec, &size);
|
||||
if (size != expected_size) {
|
||||
GGML_LOG_ERROR("Couldn't decode array from the decoder\n");
|
||||
GGML_LOG_ERROR("%s: Couldn't decode array from the decoder\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
size = 0;
|
||||
}
|
||||
@@ -322,7 +322,7 @@ static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t
|
||||
if (size) {
|
||||
val[size - 1] = '\0';
|
||||
} else {
|
||||
GGML_LOG_ERROR("Couldn't decode the blog array\n");
|
||||
GGML_LOG_ERROR("%s: Couldn't decode the blog array\n", __func__);
|
||||
apir_decoder_set_fatal(dec);
|
||||
}
|
||||
}
|
||||
@@ -332,7 +332,8 @@ static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t
|
||||
static inline void * apir_decoder_alloc_array(size_t size, size_t count) {
|
||||
size_t alloc_size;
|
||||
if (unlikely(__builtin_mul_overflow(size, count, &alloc_size))) {
|
||||
GGML_LOG_ERROR("overflow in array allocation of %zu * %zu bytes\n", size, count);
|
||||
GGML_LOG_ERROR("%s: overflow in array allocation of %zu * %zu bytes\n",
|
||||
__func__, size, count);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
|
||||
@@ -39,11 +39,17 @@ static inline void apir_encode_ggml_tensor(apir_encoder * enc, const ggml_tensor
|
||||
|
||||
static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) {
|
||||
const apir_rpc_tensor * apir_rpc_tensor = apir_decode_apir_rpc_tensor_inplace(dec);
|
||||
|
||||
if (!apir_rpc_tensor) {
|
||||
return NULL;
|
||||
}
|
||||
|
||||
ggml_init_params params{
|
||||
/*.mem_size =*/ ggml_tensor_overhead(),
|
||||
/*.mem_buffer =*/ NULL,
|
||||
/*.no_alloc =*/ true,
|
||||
};
|
||||
|
||||
ggml_context * ctx = ggml_init(params);
|
||||
|
||||
const ggml_tensor * tensor = apir_deserialize_tensor(ctx, apir_rpc_tensor);
|
||||
@@ -71,6 +77,10 @@ static inline ggml_backend_buffer_type_t apir_decode_ggml_buffer_type(apir_decod
|
||||
return (ggml_backend_buffer_type_t) handle;
|
||||
}
|
||||
|
||||
static inline void apir_encode_apir_buffer_type_host_handle(apir_encoder * enc, apir_buffer_type_host_handle_t handle) {
|
||||
apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
|
||||
}
|
||||
|
||||
static inline apir_buffer_type_host_handle_t apir_decode_apir_buffer_type_host_handle(apir_decoder * dec) {
|
||||
apir_buffer_type_host_handle_t handle;
|
||||
|
||||
@@ -154,13 +164,13 @@ static inline void apir_encode_ggml_tensor_inline(apir_encoder * enc, const ggml
|
||||
size_t tensor_size = sizeof(*tensor);
|
||||
|
||||
if (tensor->extra) {
|
||||
GGML_ABORT("Cannot pass tensors with extra");
|
||||
GGML_ABORT("%s: Cannot pass tensors with extra", __func__);
|
||||
}
|
||||
|
||||
if (tensor->src[0] && tensor->buffer) {
|
||||
static int first = 1;
|
||||
if (first) {
|
||||
GGML_LOG_WARN("Cannot pass tensors with src and buffer\n");
|
||||
GGML_LOG_WARN("%s: Cannot pass tensors with src and buffer\n", __func__);
|
||||
first = 0;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -6,7 +6,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
|
||||
|
||||
ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
|
||||
if (!context) {
|
||||
GGML_ABORT("Couldn't allocate the buffer context ...");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
|
||||
}
|
||||
|
||||
context->gpu = gpu;
|
||||
@@ -20,7 +20,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
|
||||
context->base = context->apir_context.shmem.mmap_ptr;
|
||||
context->is_from_ptr = true;
|
||||
} else {
|
||||
context->apir_context = apir_buffer_type_alloc_buffer(gpu, buft, size);
|
||||
context->apir_context = apir_buffer_type_alloc_buffer(gpu, gpu->cached_buffer_type.host_handle, size);
|
||||
context->is_from_ptr = false;
|
||||
context->base = NULL;
|
||||
}
|
||||
@@ -34,36 +34,19 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
|
||||
static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
return apir_buffer_type_get_name(gpu, buft);
|
||||
return gpu->cached_buffer_type.name;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
static size_t align = 0;
|
||||
|
||||
if (align == 0) {
|
||||
align = apir_buffer_type_get_alignment(gpu, buft);
|
||||
}
|
||||
|
||||
return align;
|
||||
return gpu->cached_buffer_type.alignment;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
static size_t max_size = 0;
|
||||
if (max_size == 0) {
|
||||
max_size = apir_buffer_type_get_max_size(gpu, buft);
|
||||
}
|
||||
|
||||
return max_size;
|
||||
}
|
||||
|
||||
static bool ggml_backend_remoting_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
|
||||
virtgpu * gpu = BUFT_TO_GPU(buft);
|
||||
|
||||
return apir_buffer_type_is_host(gpu, buft);
|
||||
return gpu->cached_buffer_type.max_size;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft,
|
||||
@@ -76,7 +59,7 @@ static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buff
|
||||
return ggml_nbytes(tensor);
|
||||
}
|
||||
|
||||
return apir_buffer_type_get_alloc_size(gpu, buft, tensor);
|
||||
return apir_buffer_type_get_alloc_size(gpu, gpu->cached_buffer_type.host_handle, tensor);
|
||||
}
|
||||
|
||||
const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface = {
|
||||
|
||||
@@ -3,32 +3,27 @@
|
||||
static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
return apir_device_get_name(gpu);
|
||||
return gpu->cached_device_info.name;
|
||||
}
|
||||
|
||||
static const char * ggml_backend_remoting_device_get_description(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
return apir_device_get_description(gpu);
|
||||
// Return the pre-cached description from the virtgpu structure
|
||||
return gpu->cached_device_info.description;
|
||||
}
|
||||
|
||||
static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
static enum ggml_backend_dev_type type;
|
||||
static bool has_type = false;
|
||||
if (!has_type) {
|
||||
has_type = true;
|
||||
type = (enum ggml_backend_dev_type) apir_device_get_type(gpu);
|
||||
}
|
||||
|
||||
return type;
|
||||
return (enum ggml_backend_dev_type) gpu->cached_device_info.type;
|
||||
}
|
||||
|
||||
static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
return apir_device_get_memory(gpu, free, total);
|
||||
*free = gpu->cached_device_info.memory_free;
|
||||
*total = gpu->cached_device_info.memory_total;
|
||||
}
|
||||
|
||||
static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
|
||||
@@ -77,13 +72,22 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_
|
||||
ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
|
||||
static std::atomic<bool> initialized = false;
|
||||
static ggml_backend_buffer_type buft;
|
||||
|
||||
static ggml_backend_buffer_type buft{
|
||||
/* .iface = */ ggml_backend_remoting_buffer_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) ctx,
|
||||
};
|
||||
if (!initialized) {
|
||||
static std::mutex mutex;
|
||||
std::lock_guard<std::mutex> lock(mutex);
|
||||
|
||||
if (!initialized) {
|
||||
buft = {
|
||||
/* .iface = */ ggml_backend_remoting_buffer_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) gpu->cached_buffer_type.host_handle,
|
||||
};
|
||||
initialized = true;
|
||||
}
|
||||
}
|
||||
|
||||
return &buft;
|
||||
}
|
||||
@@ -91,13 +95,22 @@ ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_bac
|
||||
static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) {
|
||||
virtgpu * gpu = DEV_TO_GPU(dev);
|
||||
|
||||
apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
|
||||
static std::atomic<bool> initialized = false;
|
||||
static ggml_backend_buffer_type buft;
|
||||
|
||||
static ggml_backend_buffer_type buft{
|
||||
/* .iface = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) ctx,
|
||||
};
|
||||
if (!initialized) {
|
||||
static std::mutex mutex;
|
||||
std::lock_guard<std::mutex> lock(mutex);
|
||||
|
||||
if (!initialized) {
|
||||
buft = {
|
||||
/* .iface = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
|
||||
/* .device = */ dev,
|
||||
/* .context = */ (void *) gpu->cached_buffer_type.host_handle,
|
||||
};
|
||||
initialized = true;
|
||||
}
|
||||
}
|
||||
|
||||
return &buft;
|
||||
}
|
||||
@@ -110,7 +123,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_device_buffer_from_ptr(ggml_b
|
||||
|
||||
ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
|
||||
if (!context) {
|
||||
GGML_ABORT("Couldn't allocate the buffer context ...");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the buffer context ...", __func__);
|
||||
}
|
||||
|
||||
context->gpu = gpu;
|
||||
|
||||
@@ -4,37 +4,70 @@
|
||||
#include <iostream>
|
||||
#include <mutex>
|
||||
|
||||
void ggml_virtgpu_cleanup(virtgpu * gpu);
|
||||
|
||||
static virtgpu * apir_initialize() {
|
||||
static virtgpu * apir_gpu_instance = NULL;
|
||||
static bool apir_initialized = false;
|
||||
static virtgpu * gpu = NULL;
|
||||
static std::atomic<bool> initialized = false;
|
||||
|
||||
if (initialized) {
|
||||
// fast track
|
||||
return gpu;
|
||||
}
|
||||
|
||||
{
|
||||
static std::mutex mutex;
|
||||
std::lock_guard<std::mutex> lock(mutex);
|
||||
|
||||
if (apir_initialized) {
|
||||
return apir_gpu_instance;
|
||||
if (initialized) {
|
||||
// thread safe
|
||||
return gpu;
|
||||
}
|
||||
|
||||
apir_gpu_instance = create_virtgpu();
|
||||
if (!apir_gpu_instance) {
|
||||
GGML_ABORT("failed to initialize the virtgpu");
|
||||
gpu = create_virtgpu();
|
||||
if (!gpu) {
|
||||
initialized = true;
|
||||
return NULL;
|
||||
}
|
||||
|
||||
apir_initialized = true;
|
||||
// Pre-fetch and cache all device information, it will not change
|
||||
gpu->cached_device_info.description = apir_device_get_description(gpu);
|
||||
if (!gpu->cached_device_info.description) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device description", __func__);
|
||||
}
|
||||
gpu->cached_device_info.name = apir_device_get_name(gpu);
|
||||
if (!gpu->cached_device_info.name) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu device name", __func__);
|
||||
}
|
||||
gpu->cached_device_info.device_count = apir_device_get_count(gpu);
|
||||
gpu->cached_device_info.type = apir_device_get_type(gpu);
|
||||
|
||||
apir_device_get_memory(gpu,
|
||||
&gpu->cached_device_info.memory_free,
|
||||
&gpu->cached_device_info.memory_total);
|
||||
|
||||
apir_buffer_type_host_handle_t buft_host_handle = apir_device_get_buffer_type(gpu);
|
||||
gpu->cached_buffer_type.host_handle = buft_host_handle;
|
||||
gpu->cached_buffer_type.name = apir_buffer_type_get_name(gpu, buft_host_handle);
|
||||
if (!gpu->cached_buffer_type.name) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to initialize the virtgpu buffer type name", __func__);
|
||||
}
|
||||
gpu->cached_buffer_type.alignment = apir_buffer_type_get_alignment(gpu, buft_host_handle);
|
||||
gpu->cached_buffer_type.max_size = apir_buffer_type_get_max_size(gpu, buft_host_handle);
|
||||
|
||||
initialized = true;
|
||||
}
|
||||
|
||||
return apir_gpu_instance;
|
||||
return gpu;
|
||||
}
|
||||
|
||||
static int ggml_backend_remoting_get_device_count() {
|
||||
virtgpu * gpu = apir_initialize();
|
||||
if (!gpu) {
|
||||
GGML_LOG_WARN("apir_initialize failed\n");
|
||||
return 0;
|
||||
}
|
||||
|
||||
return apir_device_get_count(gpu);
|
||||
return gpu->cached_device_info.device_count;
|
||||
}
|
||||
|
||||
static size_t ggml_backend_remoting_reg_get_device_count(ggml_backend_reg_t reg) {
|
||||
@@ -52,17 +85,21 @@ ggml_backend_dev_t ggml_backend_remoting_get_device(size_t device) {
|
||||
|
||||
static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
|
||||
if (devices.size() > 0) {
|
||||
GGML_LOG_INFO("%s: already initialized\n", __func__);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "%s: already initialized\n", __func__);
|
||||
return;
|
||||
}
|
||||
|
||||
virtgpu * gpu = apir_initialize();
|
||||
if (!gpu) {
|
||||
GGML_LOG_ERROR("apir_initialize failed\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: apir_initialize failed\n", __func__);
|
||||
return;
|
||||
}
|
||||
|
||||
static bool initialized = false;
|
||||
static std::atomic<bool> initialized = false;
|
||||
|
||||
if (initialized) {
|
||||
return; // fast track
|
||||
}
|
||||
|
||||
{
|
||||
static std::mutex mutex;
|
||||
@@ -70,10 +107,10 @@ static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
|
||||
if (!initialized) {
|
||||
for (int i = 0; i < ggml_backend_remoting_get_device_count(); i++) {
|
||||
ggml_backend_remoting_device_context * ctx = new ggml_backend_remoting_device_context;
|
||||
char desc[256] = "API Remoting device";
|
||||
char desc[256] = "ggml-virtgpu API Remoting device";
|
||||
|
||||
ctx->device = i;
|
||||
ctx->name = GGML_REMOTING_FRONTEND_NAME + std::to_string(i);
|
||||
ctx->name = GGML_VIRTGPU_NAME + std::to_string(i);
|
||||
ctx->description = desc;
|
||||
ctx->gpu = gpu;
|
||||
|
||||
@@ -98,7 +135,7 @@ static ggml_backend_dev_t ggml_backend_remoting_reg_get_device(ggml_backend_reg_
|
||||
static const char * ggml_backend_remoting_reg_get_name(ggml_backend_reg_t reg) {
|
||||
UNUSED(reg);
|
||||
|
||||
return GGML_REMOTING_FRONTEND_NAME;
|
||||
return GGML_VIRTGPU_NAME;
|
||||
}
|
||||
|
||||
static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
|
||||
@@ -111,8 +148,7 @@ static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
|
||||
ggml_backend_reg_t ggml_backend_virtgpu_reg() {
|
||||
virtgpu * gpu = apir_initialize();
|
||||
if (!gpu) {
|
||||
GGML_LOG_ERROR("virtgpu_apir_initialize failed\n");
|
||||
return NULL;
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: virtgpu_apir_initialize failed\n", __func__);
|
||||
}
|
||||
|
||||
static ggml_backend_reg reg = {
|
||||
@@ -129,9 +165,25 @@ ggml_backend_reg_t ggml_backend_virtgpu_reg() {
|
||||
|
||||
ggml_backend_remoting_reg_init_devices(®);
|
||||
|
||||
GGML_LOG_INFO("%s: initialized\n", __func__);
|
||||
|
||||
return ®
|
||||
}
|
||||
|
||||
// public function, not exposed in the GGML interface at the moment
|
||||
void ggml_virtgpu_cleanup(virtgpu * gpu) {
|
||||
if (gpu->cached_device_info.name) {
|
||||
free(gpu->cached_device_info.name);
|
||||
gpu->cached_device_info.name = NULL;
|
||||
}
|
||||
if (gpu->cached_device_info.description) {
|
||||
free(gpu->cached_device_info.description);
|
||||
gpu->cached_device_info.description = NULL;
|
||||
}
|
||||
if (gpu->cached_buffer_type.name) {
|
||||
free(gpu->cached_buffer_type.name);
|
||||
gpu->cached_buffer_type.name = NULL;
|
||||
}
|
||||
|
||||
mtx_destroy(&gpu->data_shmem_mutex);
|
||||
}
|
||||
|
||||
GGML_BACKEND_DL_IMPL(ggml_backend_virtgpu_reg)
|
||||
|
||||
@@ -34,7 +34,11 @@ static ggml_backend_i ggml_backend_remoting_interface = {
|
||||
/* .free = */ ggml_backend_remoting_free,
|
||||
/* .set_tensor_async = */ NULL, // ggml_backend_remoting_set_tensor_async,
|
||||
/* .get_tensor_async = */ NULL, // ggml_backend_remoting_get_tensor_async,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL, // ggml_backend_remoting_cpy_tensor_async,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ NULL, // ggml_backend_remoting_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -8,6 +8,9 @@
|
||||
#include <memory>
|
||||
#include <string>
|
||||
|
||||
#define GGML_VIRTGPU_NAME "ggml-virtgpu"
|
||||
#define GGML_VIRTGPU "ggml-virtgpu: "
|
||||
|
||||
// USE_ALWAYS_TRUE_SUPPORTS_OP: 1 is fast, 0 avoid micro-benchmark crashes
|
||||
|
||||
#define USE_ALWAYS_TRUE_SUPPORTS_OP 1
|
||||
@@ -62,7 +65,7 @@ static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggm
|
||||
|
||||
static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) {
|
||||
if (!buffer->context) {
|
||||
GGML_ABORT("%s: no context available :/", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: no context available :/", __func__);
|
||||
}
|
||||
return BUFFER_TO_HOST_HANDLE(buffer);
|
||||
}
|
||||
|
||||
@@ -24,10 +24,10 @@ functions:
|
||||
frontend_return: "int"
|
||||
|
||||
get_name:
|
||||
frontend_return: "const char *"
|
||||
frontend_return: "char *"
|
||||
|
||||
get_description:
|
||||
frontend_return: "const char *"
|
||||
frontend_return: "char *"
|
||||
|
||||
get_type:
|
||||
frontend_return: "uint32_t"
|
||||
@@ -64,35 +64,33 @@ functions:
|
||||
group_description: "buffer-type"
|
||||
functions:
|
||||
get_name:
|
||||
frontend_return: "const char *"
|
||||
frontend_return: "char *"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
|
||||
get_alignment:
|
||||
frontend_return: "size_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
|
||||
get_max_size:
|
||||
frontend_return: "size_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
|
||||
is_host:
|
||||
frontend_return: "bool"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
deprecated: true
|
||||
|
||||
alloc_buffer:
|
||||
frontend_return: "apir_buffer_context_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buffer_buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
- "size_t size"
|
||||
|
||||
get_alloc_size:
|
||||
frontend_return: "size_t"
|
||||
frontend_extra_params:
|
||||
- "ggml_backend_buffer_type_t buft"
|
||||
- "apir_buffer_type_host_handle_t host_handle"
|
||||
- "const ggml_tensor *op"
|
||||
|
||||
buffer:
|
||||
|
||||
@@ -116,7 +116,7 @@ class RemotingCodebaseGenerator:
|
||||
'frontend_return': func_metadata.get('frontend_return', 'void'),
|
||||
'frontend_extra_params': func_metadata.get('frontend_extra_params', []),
|
||||
'group_description': group_description,
|
||||
'newly_added': func_metadata.get('newly_added', False)
|
||||
'deprecated': func_metadata.get('deprecated', False),
|
||||
})
|
||||
enum_value += 1
|
||||
|
||||
@@ -165,6 +165,9 @@ class RemotingCodebaseGenerator:
|
||||
|
||||
signature = "uint32_t"
|
||||
params = "apir_encoder *enc, apir_decoder *dec, virgl_apir_context *ctx"
|
||||
if func['deprecated']:
|
||||
decl_lines.append(f"/* {func['enum_name']} is deprecated. Keeping the handler for backward compatibility. */")
|
||||
|
||||
decl_lines.append(f"{signature} {func['backend_function']}({params});")
|
||||
|
||||
# Switch cases
|
||||
@@ -176,7 +179,9 @@ class RemotingCodebaseGenerator:
|
||||
switch_lines.append(f" /* {func['group_description']} */")
|
||||
current_group = func['group_name']
|
||||
|
||||
switch_lines.append(f" case {func['enum_name']}: return \"{func['backend_function']}\";")
|
||||
deprecated = " (DEPRECATED)" if func['deprecated'] else ""
|
||||
|
||||
switch_lines.append(f" case {func['enum_name']}: return \"{func['backend_function']}{deprecated}\";")
|
||||
|
||||
# Dispatch table
|
||||
table_lines = []
|
||||
@@ -188,7 +193,8 @@ class RemotingCodebaseGenerator:
|
||||
table_lines.append("")
|
||||
current_group = func['group_name']
|
||||
|
||||
table_lines.append(f" /* {func['enum_name']} = */ {func['backend_function']},")
|
||||
deprecated = " /* DEPRECATED */" if func['deprecated'] else ""
|
||||
table_lines.append(f" /* {func['enum_name']} = */ {func['backend_function']}{deprecated},")
|
||||
|
||||
header_content = f'''\
|
||||
#pragma once
|
||||
@@ -225,6 +231,10 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
|
||||
decl_lines.append(f"/* {func['group_description']} */")
|
||||
current_group = func['group_name']
|
||||
|
||||
if func['deprecated']:
|
||||
decl_lines.append(f"/* {func['frontend_function']} is deprecated. */")
|
||||
continue
|
||||
|
||||
# Build parameter list
|
||||
params = [self.naming_patterns['frontend_base_param']]
|
||||
params.extend(func['frontend_extra_params'])
|
||||
@@ -287,7 +297,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
|
||||
generated_files = [apir_backend_path, backend_dispatched_path, virtgpu_forward_path]
|
||||
|
||||
if not self.clang_format_available:
|
||||
logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted."
|
||||
logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted.\n"
|
||||
" Install clang-format to enable automatic code formatting.")
|
||||
else:
|
||||
logging.info("\n🎨 Formatting files with clang-format...")
|
||||
|
||||
@@ -18,12 +18,17 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
|
||||
|
||||
virtgpu_shmem temp_shmem; // Local storage for large buffers
|
||||
virtgpu_shmem * shmem = &temp_shmem;
|
||||
bool using_shared_shmem = false;
|
||||
|
||||
if (cgraph_size <= gpu->data_shmem.mmap_size) {
|
||||
// prefer the init-time allocated page, if large enough
|
||||
// Lock mutex before using shared data_shmem buffer
|
||||
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
|
||||
}
|
||||
using_shared_shmem = true;
|
||||
shmem = &gpu->data_shmem;
|
||||
} else if (virtgpu_shmem_create(gpu, cgraph_size, shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
|
||||
}
|
||||
|
||||
apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
|
||||
@@ -42,7 +47,10 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
if (shmem != &gpu->data_shmem) {
|
||||
// Unlock mutex before cleanup
|
||||
if (using_shared_shmem) {
|
||||
mtx_unlock(&gpu->data_shmem_mutex);
|
||||
} else {
|
||||
virtgpu_shmem_destroy(gpu, shmem);
|
||||
}
|
||||
|
||||
|
||||
@@ -1,20 +1,20 @@
|
||||
#include "virtgpu-forward-impl.h"
|
||||
|
||||
const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
char * apir_buffer_type_get_name(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
const size_t string_size = apir_decode_array_size_unchecked(decoder);
|
||||
char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
if (!string) {
|
||||
GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device name buffer\n", __func__);
|
||||
apir_decoder_set_fatal(decoder);
|
||||
}
|
||||
apir_decode_char_array(decoder, string, string_size);
|
||||
@@ -24,14 +24,14 @@ const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t
|
||||
return string;
|
||||
}
|
||||
|
||||
size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
size_t apir_buffer_type_get_alignment(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
@@ -43,14 +43,14 @@ size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t
|
||||
return alignment;
|
||||
}
|
||||
|
||||
size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
size_t apir_buffer_type_get_max_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
@@ -62,26 +62,7 @@ size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t b
|
||||
return max_size;
|
||||
}
|
||||
|
||||
bool apir_buffer_type_is_host(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
bool is_host;
|
||||
apir_decode_bool_t(decoder, &is_host);
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
return is_host;
|
||||
}
|
||||
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_buffer_type_t buft, size_t size) {
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, size_t size) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -90,7 +71,7 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
apir_encode_size_t(encoder, &size);
|
||||
|
||||
@@ -103,14 +84,14 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
|
||||
return buffer_context;
|
||||
}
|
||||
|
||||
size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op) {
|
||||
size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, const ggml_tensor * op) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE);
|
||||
|
||||
apir_encode_ggml_buffer_type(encoder, buft);
|
||||
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
|
||||
|
||||
apir_encode_ggml_tensor_inline(encoder, op);
|
||||
|
||||
|
||||
@@ -36,13 +36,18 @@ void apir_buffer_set_tensor(virtgpu * gpu,
|
||||
|
||||
virtgpu_shmem temp_shmem; // Local storage for large buffers
|
||||
virtgpu_shmem * shmem = &temp_shmem;
|
||||
bool using_shared_shmem = false;
|
||||
|
||||
if (size <= gpu->data_shmem.mmap_size) {
|
||||
// prefer the init-time allocated page, if large enough
|
||||
// Lock mutex before using shared data_shmem buffer
|
||||
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
|
||||
}
|
||||
using_shared_shmem = true;
|
||||
shmem = &gpu->data_shmem;
|
||||
|
||||
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
|
||||
}
|
||||
|
||||
memcpy(shmem->mmap_ptr, data, size);
|
||||
@@ -55,7 +60,10 @@ void apir_buffer_set_tensor(virtgpu * gpu,
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
if (shmem != &gpu->data_shmem) {
|
||||
// Unlock mutex before cleanup
|
||||
if (using_shared_shmem) {
|
||||
mtx_unlock(&gpu->data_shmem_mutex);
|
||||
} else {
|
||||
virtgpu_shmem_destroy(gpu, shmem);
|
||||
}
|
||||
|
||||
@@ -79,13 +87,18 @@ void apir_buffer_get_tensor(virtgpu * gpu,
|
||||
|
||||
virtgpu_shmem temp_shmem; // Local storage for large buffers
|
||||
virtgpu_shmem * shmem = &temp_shmem;
|
||||
bool using_shared_shmem = false;
|
||||
|
||||
if (size <= gpu->data_shmem.mmap_size) {
|
||||
// prefer the init-time allocated page, if large enough
|
||||
// Lock mutex before using shared data_shmem buffer
|
||||
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
|
||||
}
|
||||
using_shared_shmem = true;
|
||||
shmem = &gpu->data_shmem;
|
||||
|
||||
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
|
||||
}
|
||||
|
||||
apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
|
||||
@@ -98,7 +111,10 @@ void apir_buffer_get_tensor(virtgpu * gpu,
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
if (shmem != &gpu->data_shmem) {
|
||||
// Unlock mutex before cleanup
|
||||
if (using_shared_shmem) {
|
||||
mtx_unlock(&gpu->data_shmem_mutex);
|
||||
} else {
|
||||
virtgpu_shmem_destroy(gpu, shmem);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -2,11 +2,6 @@
|
||||
#include "virtgpu-shm.h"
|
||||
|
||||
int apir_device_get_count(virtgpu * gpu) {
|
||||
static int32_t dev_count = -1;
|
||||
if (dev_count != -1) {
|
||||
return dev_count;
|
||||
}
|
||||
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -14,6 +9,7 @@ int apir_device_get_count(virtgpu * gpu) {
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_COUNT);
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
int32_t dev_count = -1;
|
||||
apir_decode_int32_t(decoder, &dev_count);
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
@@ -21,11 +17,7 @@ int apir_device_get_count(virtgpu * gpu) {
|
||||
return dev_count;
|
||||
}
|
||||
|
||||
const char * apir_device_get_name(virtgpu * gpu) {
|
||||
static char * string = nullptr;
|
||||
if (string) {
|
||||
return string;
|
||||
}
|
||||
char * apir_device_get_name(virtgpu * gpu) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -34,9 +26,9 @@ const char * apir_device_get_name(virtgpu * gpu) {
|
||||
REMOTE_CALL(gpu, encoder, decoder, ret);
|
||||
|
||||
const size_t string_size = apir_decode_array_size_unchecked(decoder);
|
||||
string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
if (!string) {
|
||||
GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device name buffer\n", __func__);
|
||||
return NULL;
|
||||
}
|
||||
apir_decode_char_array(decoder, string, string_size);
|
||||
@@ -46,7 +38,7 @@ const char * apir_device_get_name(virtgpu * gpu) {
|
||||
return string;
|
||||
}
|
||||
|
||||
const char * apir_device_get_description(virtgpu * gpu) {
|
||||
char * apir_device_get_description(virtgpu * gpu) {
|
||||
apir_encoder * encoder;
|
||||
apir_decoder * decoder;
|
||||
ApirForwardReturnCode ret;
|
||||
@@ -58,7 +50,7 @@ const char * apir_device_get_description(virtgpu * gpu) {
|
||||
const size_t string_size = apir_decode_array_size_unchecked(decoder);
|
||||
char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
|
||||
if (!string) {
|
||||
GGML_LOG_ERROR("%s: Could not allocate the device description buffer\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Could not allocate the device description buffer\n", __func__);
|
||||
|
||||
return NULL;
|
||||
}
|
||||
@@ -181,7 +173,7 @@ apir_buffer_context_t apir_device_buffer_from_ptr(virtgpu * gpu, size_t size, si
|
||||
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR);
|
||||
|
||||
if (virtgpu_shmem_create(gpu, size, &buffer_context.shmem)) {
|
||||
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
|
||||
GGML_ABORT(GGML_VIRTGPU "Couldn't allocate the guest-host shared buffer");
|
||||
}
|
||||
|
||||
apir_encode_virtgpu_shmem_res_id(encoder, buffer_context.shmem.res_id);
|
||||
|
||||
@@ -11,7 +11,7 @@
|
||||
int32_t forward_flag = (int32_t) apir_command_type__; \
|
||||
encoder_name = remote_call_prepare(gpu_dev_name, APIR_COMMAND_TYPE_FORWARD, forward_flag); \
|
||||
if (!encoder_name) { \
|
||||
GGML_ABORT("%s: failed to prepare the remote call encoder", __func__); \
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to prepare the remote call encoder", __func__); \
|
||||
} \
|
||||
} while (0)
|
||||
|
||||
@@ -19,10 +19,10 @@
|
||||
do { \
|
||||
ret_name = (ApirForwardReturnCode) remote_call(gpu_dev_name, encoder_name, &decoder_name, 0, NULL); \
|
||||
if (!decoder_name) { \
|
||||
GGML_ABORT("%s: failed to kick the remote call", __func__); \
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to kick the remote call", __func__); \
|
||||
} \
|
||||
if (ret_name < APIR_FORWARD_BASE_INDEX) { \
|
||||
GGML_ABORT("%s: failed to forward the API call: %s: code %d", __func__, \
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to forward the API call: %s: code %d", __func__, \
|
||||
apir_forward_error(ret_name), ret_name); \
|
||||
} \
|
||||
ret_name = (ApirForwardReturnCode) (ret_name - APIR_FORWARD_BASE_INDEX); \
|
||||
|
||||
@@ -3,8 +3,8 @@
|
||||
/* device */
|
||||
void apir_device_get_device_count(struct virtgpu * gpu);
|
||||
int apir_device_get_count(struct virtgpu * gpu);
|
||||
const char * apir_device_get_name(struct virtgpu * gpu);
|
||||
const char * apir_device_get_description(struct virtgpu * gpu);
|
||||
char * apir_device_get_name(struct virtgpu * gpu);
|
||||
char * apir_device_get_description(struct virtgpu * gpu);
|
||||
uint32_t apir_device_get_type(struct virtgpu * gpu);
|
||||
void apir_device_get_memory(struct virtgpu * gpu, size_t * free, size_t * total);
|
||||
bool apir_device_supports_op(struct virtgpu * gpu, const ggml_tensor * op);
|
||||
@@ -17,14 +17,15 @@ void apir_device_get_props(struct virtgpu * gpu,
|
||||
apir_buffer_context_t apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);
|
||||
|
||||
/* buffer-type */
|
||||
const char * apir_buffer_type_get_name(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
size_t apir_buffer_type_get_alignment(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
size_t apir_buffer_type_get_max_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
bool apir_buffer_type_is_host(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu * gpu,
|
||||
ggml_backend_buffer_type_t buffer_buft,
|
||||
size_t size);
|
||||
size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op);
|
||||
char * apir_buffer_type_get_name(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
|
||||
size_t apir_buffer_type_get_alignment(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
|
||||
size_t apir_buffer_type_get_max_size(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
|
||||
apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu * gpu,
|
||||
apir_buffer_type_host_handle_t host_handle,
|
||||
size_t size);
|
||||
size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu,
|
||||
apir_buffer_type_host_handle_t host_handle,
|
||||
const ggml_tensor * op);
|
||||
|
||||
/* buffer */
|
||||
void * apir_buffer_get_base(struct virtgpu * gpu, apir_buffer_context_t * buffer_context);
|
||||
|
||||
@@ -85,8 +85,7 @@ int virtgpu_shmem_create(virtgpu * gpu, size_t size, virtgpu_shmem * shmem) {
|
||||
void * ptr = virtgpu_ioctl_map(gpu, gem_handle, size);
|
||||
if (!ptr) {
|
||||
virtgpu_ioctl_gem_close(gpu, gem_handle);
|
||||
GGML_LOG_ERROR("virtgpu_ioctl_map FAILED\n");
|
||||
exit(1);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: virtgpu_ioctl_map failed\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
|
||||
@@ -33,7 +33,7 @@ static int virtgpu_handshake(virtgpu * gpu) {
|
||||
|
||||
encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_HANDSHAKE, 0);
|
||||
if (!encoder) {
|
||||
GGML_ABORT("%s: failed to prepare the remote call encoder", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: failed to prepare the remote call encoder", __func__);
|
||||
return 1;
|
||||
}
|
||||
|
||||
@@ -52,7 +52,7 @@ static int virtgpu_handshake(virtgpu * gpu) {
|
||||
log_call_duration(call_duration_ns, "API Remoting handshake");
|
||||
|
||||
if (!decoder) {
|
||||
GGML_ABORT(
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to initiate the communication with the virglrenderer library. "
|
||||
"Most likely, the wrong virglrenderer library was loaded in the hypervisor.",
|
||||
__func__);
|
||||
@@ -65,7 +65,8 @@ static int virtgpu_handshake(virtgpu * gpu) {
|
||||
uint32_t host_minor;
|
||||
|
||||
if (ret_magic != APIR_HANDSHAKE_MAGIC) {
|
||||
GGML_ABORT("%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic,
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic,
|
||||
apir_backend_initialize_error(ret_magic));
|
||||
} else {
|
||||
apir_decode_uint32_t(decoder, &host_major);
|
||||
@@ -78,13 +79,13 @@ static int virtgpu_handshake(virtgpu * gpu) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
GGML_LOG_INFO("%s: Guest is running with %u.%u\n", __func__, guest_major, guest_minor);
|
||||
GGML_LOG_INFO("%s: Host is running with %u.%u\n", __func__, host_major, host_minor);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "%s: Guest is running with %u.%u\n", __func__, guest_major, guest_minor);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "%s: Host is running with %u.%u\n", __func__, host_major, host_minor);
|
||||
|
||||
if (guest_major != host_major) {
|
||||
GGML_LOG_ERROR("Host major (%d) and guest major (%d) version differ\n", host_major, guest_major);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "Host major (%d) and guest major (%d) version differ\n", host_major, guest_major);
|
||||
} else if (guest_minor != host_minor) {
|
||||
GGML_LOG_WARN("Host minor (%d) and guest minor (%d) version differ\n", host_minor, guest_minor);
|
||||
GGML_LOG_WARN(GGML_VIRTGPU "Host minor (%d) and guest minor (%d) version differ\n", host_minor, guest_minor);
|
||||
}
|
||||
|
||||
return 0;
|
||||
@@ -97,7 +98,7 @@ static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
|
||||
|
||||
encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_LOADLIBRARY, 0);
|
||||
if (!encoder) {
|
||||
GGML_ABORT("%s: hypercall error: failed to prepare the remote call encoder", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: hypercall error: failed to prepare the API Remoting command encoder", __func__);
|
||||
return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR;
|
||||
}
|
||||
|
||||
@@ -108,36 +109,67 @@ static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
|
||||
log_call_duration(call_duration_ns, "API Remoting LoadLibrary");
|
||||
|
||||
if (!decoder) {
|
||||
GGML_ABORT("%s: hypercall error: failed to kick the API remoting hypercall.\n", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: hypercall error: failed to trigger the API Remoting hypercall.\n", __func__);
|
||||
return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR;
|
||||
}
|
||||
|
||||
remote_call_finish(gpu, encoder, decoder);
|
||||
|
||||
if (ret == APIR_LOAD_LIBRARY_SUCCESS) {
|
||||
GGML_LOG_INFO("%s: The API Remoting backend was successfully loaded and initialized\n", __func__);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "The API Remoting backend was successfully loaded and initialized\n");
|
||||
|
||||
return ret;
|
||||
}
|
||||
|
||||
// something wrong happened, find out what.
|
||||
|
||||
if (ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
|
||||
GGML_ABORT("%s: virglrenderer could not load the API Remoting backend library: %s (code %d)", __func__,
|
||||
apir_load_library_error(ret), ret);
|
||||
if (ret == APIR_LOAD_LIBRARY_ENV_VAR_MISSING) {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: virglrenderer could not open the API Remoting backend library, "
|
||||
"some environment variables are missing. "
|
||||
"Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
|
||||
__func__, apir_load_library_error(ret));
|
||||
} else if (ret == APIR_LOAD_LIBRARY_CANNOT_OPEN) {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: virglrenderer could not open the API Remoting backend library. "
|
||||
"Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
|
||||
__func__, apir_load_library_error(ret));
|
||||
} else if (ret == APIR_LOAD_LIBRARY_ENV_VAR_MISSING) {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: could not load the backend library, some symbols are missing. "
|
||||
"Make sure virglrenderer is correctly configured by the hypervisor. (%s) ",
|
||||
__func__, apir_load_library_error(ret));
|
||||
} else {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: virglrenderer could not load the API Remoting backend library. (%s - code %d)", __func__,
|
||||
apir_load_library_error(ret), ret);
|
||||
}
|
||||
return ret;
|
||||
}
|
||||
|
||||
GGML_LOG_INFO("%s: virglrenderer successfully loaded the API Remoting backend library", __func__);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU
|
||||
"%s: virglrenderer successfully loaded the API Remoting backend library.\n", __func__);
|
||||
|
||||
ApirLoadLibraryReturnCode apir_ret = (ApirLoadLibraryReturnCode) (ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX);
|
||||
|
||||
if (apir_ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
|
||||
GGML_ABORT("%s: the API Remoting backend library couldn't load the backend library: apir code=%d | %s)",
|
||||
if (apir_ret == APIR_LOAD_LIBRARY_CANNOT_OPEN) {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: the API Remoting backend library couldn't load the GGML backend library. "
|
||||
"Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
|
||||
__func__, apir_load_library_error(apir_ret));
|
||||
} else if (apir_ret == APIR_LOAD_LIBRARY_SYMBOL_MISSING) {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: the API Remoting backend library couldn't load the GGML backend library, some symbols are missing. "
|
||||
"Make sure virglrenderer is correctly configured by the hypervisor. (%s)",
|
||||
__func__, apir_load_library_error(apir_ret));
|
||||
} else if (apir_ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: the API Remoting backend library couldn't load the GGML backend library: apir code=%d | %s)",
|
||||
__func__, apir_ret, apir_load_library_error(apir_ret));
|
||||
} else {
|
||||
uint32_t lib_ret = apir_ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX;
|
||||
GGML_ABORT("%s: the API Remoting backend library initialize its backend library: apir code=%d)", __func__,
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: the API Remoting backend library initialize its backend library: apir code=%d)", __func__,
|
||||
lib_ret);
|
||||
}
|
||||
return ret;
|
||||
@@ -149,38 +181,58 @@ virtgpu * create_virtgpu() {
|
||||
gpu->use_apir_capset = getenv("GGML_REMOTING_USE_APIR_CAPSET") != nullptr;
|
||||
util_sparse_array_init(&gpu->shmem_array, sizeof(virtgpu_shmem), 1024);
|
||||
|
||||
// Initialize mutex to protect shared data_shmem buffer
|
||||
if (mtx_init(&gpu->data_shmem_mutex, mtx_plain) != thrd_success) {
|
||||
delete gpu;
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to initialize data_shmem mutex", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (virtgpu_open(gpu) != APIR_SUCCESS) {
|
||||
GGML_ABORT("%s: failed to open the virtgpu device", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU
|
||||
"%s: failed to open the virtgpu device\n", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (virtgpu_init_capset(gpu) != APIR_SUCCESS) {
|
||||
GGML_ABORT("%s: failed to initialize the GPU capset", __func__);
|
||||
if (gpu->use_apir_capset) {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to initialize the virtgpu APIR capset. Make sure that the virglrenderer library supports it.", __func__);
|
||||
} else {
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to initialize the virtgpu Venus capset", __func__);
|
||||
}
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (virtgpu_init_context(gpu) != APIR_SUCCESS) {
|
||||
GGML_ABORT("%s: failed to initialize the GPU context", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to initialize the GPU context", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (virtgpu_shmem_create(gpu, SHMEM_REPLY_SIZE, &gpu->reply_shmem)) {
|
||||
GGML_ABORT("%s: failed to create the shared reply memory pages", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to create the shared reply memory pages", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (virtgpu_shmem_create(gpu, SHMEM_DATA_SIZE, &gpu->data_shmem)) {
|
||||
GGML_ABORT("%s: failed to create the shared data memory pages", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to create the shared data memory pages", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (virtgpu_handshake(gpu)) {
|
||||
GGML_ABORT("%s: failed to handshake with the virglrenderer library", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to handshake with the virglrenderer library", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (virtgpu_load_library(gpu) != APIR_LOAD_LIBRARY_SUCCESS) {
|
||||
GGML_ABORT("%s: failed to load the backend library", __func__);
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to load the backend library", __func__);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@@ -191,7 +243,8 @@ static virt_gpu_result_t virtgpu_open(virtgpu * gpu) {
|
||||
drmDevicePtr devs[8];
|
||||
int count = drmGetDevices2(0, devs, ARRAY_SIZE(devs));
|
||||
if (count < 0) {
|
||||
GGML_LOG_ERROR("%s: failed to enumerate DRM devices\n", __func__);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU
|
||||
"%s: failed to enumerate DRM devices\n", __func__);
|
||||
return APIR_ERROR_INITIALIZATION_FAILED;
|
||||
}
|
||||
|
||||
@@ -213,16 +266,19 @@ static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr d
|
||||
|
||||
int fd = open(node_path, O_RDWR | O_CLOEXEC);
|
||||
if (fd < 0) {
|
||||
GGML_ABORT("failed to open %s", node_path);
|
||||
GGML_ABORT(GGML_VIRTGPU
|
||||
"%s: failed to open %s", __func__, node_path);
|
||||
return APIR_ERROR_INITIALIZATION_FAILED;
|
||||
}
|
||||
|
||||
drmVersionPtr version = drmGetVersion(fd);
|
||||
if (!version || strcmp(version->name, "virtio_gpu") || version->version_major != 0) {
|
||||
if (version) {
|
||||
GGML_ABORT("unknown DRM driver %s version %d", version->name, version->version_major);
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU
|
||||
"%s: unknown DRM driver %s version %d\n", __func__, version->name, version->version_major);
|
||||
} else {
|
||||
GGML_ABORT("failed to get DRM driver version");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU
|
||||
"%s: failed to get DRM driver version\n", __func__);
|
||||
}
|
||||
|
||||
if (version) {
|
||||
@@ -236,7 +292,7 @@ static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr d
|
||||
|
||||
drmFreeVersion(version);
|
||||
|
||||
GGML_LOG_INFO("using DRM device %s\n", node_path);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "using DRM device %s\n", node_path);
|
||||
|
||||
return APIR_SUCCESS;
|
||||
}
|
||||
@@ -245,7 +301,7 @@ static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu) {
|
||||
assert(!gpu->capset.version);
|
||||
const int ret = virtgpu_ioctl_context_init(gpu, gpu->capset.id);
|
||||
if (ret) {
|
||||
GGML_LOG_INFO("failed to initialize context: %s\n", strerror(errno));
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: failed to initialize context: %s\n", __func__, strerror(errno));
|
||||
return APIR_ERROR_INITIALIZATION_FAILED;
|
||||
}
|
||||
|
||||
@@ -254,10 +310,10 @@ static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu) {
|
||||
|
||||
static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu) {
|
||||
if (gpu->use_apir_capset) {
|
||||
GGML_LOG_INFO("Using the APIR capset\n");
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "Using the APIR capset\n");
|
||||
gpu->capset.id = VIRTGPU_DRM_CAPSET_APIR;
|
||||
} else {
|
||||
GGML_LOG_INFO("Using the Venus capset\n");
|
||||
GGML_LOG_INFO(GGML_VIRTGPU "Using the Venus capset\n");
|
||||
gpu->capset.id = VIRTGPU_DRM_CAPSET_VENUS;
|
||||
}
|
||||
gpu->capset.version = 0;
|
||||
@@ -266,7 +322,9 @@ static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu) {
|
||||
virtgpu_ioctl_get_caps(gpu, gpu->capset.id, gpu->capset.version, &gpu->capset.data, sizeof(gpu->capset.data));
|
||||
|
||||
if (ret) {
|
||||
GGML_LOG_INFO("failed to get APIR v%d capset: %s\n", gpu->capset.version, strerror(errno));
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU
|
||||
"%s: failed to get APIR v%d capset: %s\n",
|
||||
__func__, gpu->capset.version, strerror(errno));
|
||||
return APIR_ERROR_INITIALIZATION_FAILED;
|
||||
}
|
||||
|
||||
@@ -333,9 +391,9 @@ apir_encoder * remote_call_prepare(virtgpu * gpu, ApirCommandType apir_cmd_type,
|
||||
* Prepare the command encoder and its buffer
|
||||
*/
|
||||
|
||||
static char encoder_buffer[4096];
|
||||
thread_local char encoder_buffer[4096];
|
||||
|
||||
static apir_encoder enc;
|
||||
thread_local apir_encoder enc;
|
||||
enc = {
|
||||
.cur = encoder_buffer,
|
||||
.start = encoder_buffer,
|
||||
@@ -369,19 +427,19 @@ void remote_call_finish(virtgpu * gpu, apir_encoder * enc, apir_decoder * dec) {
|
||||
UNUSED(gpu);
|
||||
|
||||
if (!enc) {
|
||||
GGML_LOG_ERROR("Invalid (null) encoder\n");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Invalid (null) encoder", __func__);
|
||||
}
|
||||
|
||||
if (!dec) {
|
||||
GGML_LOG_ERROR("Invalid (null) decoder\n");
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: Invalid (null) decoder", __func__);
|
||||
}
|
||||
|
||||
if (apir_encoder_get_fatal(enc)) {
|
||||
GGML_LOG_ERROR("Failed to encode the output parameters.\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Failed to encode the output parameters.", __func__);
|
||||
}
|
||||
|
||||
if (apir_decoder_get_fatal(dec)) {
|
||||
GGML_LOG_ERROR("Failed to decode the input parameters.\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: Failed to decode the input parameters.", __func__);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -423,7 +481,7 @@ uint32_t remote_call(virtgpu * gpu,
|
||||
int ret = drmIoctl(gpu->fd, DRM_IOCTL_VIRTGPU_EXECBUFFER, &args);
|
||||
|
||||
if (ret != 0) {
|
||||
GGML_ABORT("%s: the virtgpu EXECBUFFER ioctl failed (%d)", __func__, ret);
|
||||
GGML_ABORT(GGML_VIRTGPU "%s: the virtgpu EXECBUFFER ioctl failed (%d)", __func__, ret);
|
||||
}
|
||||
|
||||
/*
|
||||
@@ -467,7 +525,7 @@ uint32_t remote_call(virtgpu * gpu,
|
||||
}
|
||||
|
||||
if (max_wait_ms && timedout) {
|
||||
GGML_LOG_ERROR("timed out waiting for the host answer...\n");
|
||||
GGML_LOG_ERROR(GGML_VIRTGPU "%s: timed out waiting for the host answer...\n", __func__);
|
||||
return APIR_FORWARD_TIMEOUT;
|
||||
}
|
||||
|
||||
@@ -489,10 +547,13 @@ static void log_call_duration(long long call_duration_ns, const char * name) {
|
||||
double call_duration_s = (double) call_duration_ns / 1e9; // 1 second = 1e9 nanoseconds
|
||||
|
||||
if (call_duration_s > 1) {
|
||||
GGML_LOG_INFO("%s: waited %.2fs for the %s host reply...\n", __func__, call_duration_s, name);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU
|
||||
"waited %.2fs for the %s host reply...\n", call_duration_s, name);
|
||||
} else if (call_duration_ms > 1) {
|
||||
GGML_LOG_INFO("%s: waited %.2fms for the %s host reply...\n", __func__, call_duration_ms, name);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU
|
||||
"waited %.2fms for the %s host reply...\n", call_duration_ms, name);
|
||||
} else {
|
||||
GGML_LOG_INFO("%s: waited %lldns for the %s host reply...\n", __func__, call_duration_ns, name);
|
||||
GGML_LOG_INFO(GGML_VIRTGPU
|
||||
"waited %lldns for the %s host reply...\n", call_duration_ns, name);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -17,6 +17,8 @@
|
||||
|
||||
#include <cstring>
|
||||
|
||||
#include "ggml-remoting.h"
|
||||
|
||||
#define VIRGL_RENDERER_UNSTABLE_APIS 1
|
||||
#include "apir_hw.h"
|
||||
#include <drm/virtgpu_drm.h>
|
||||
@@ -73,6 +75,27 @@ struct virtgpu {
|
||||
/* APIR communication pages */
|
||||
virtgpu_shmem reply_shmem;
|
||||
virtgpu_shmem data_shmem;
|
||||
|
||||
/* Mutex to protect shared data_shmem buffer from concurrent access */
|
||||
mtx_t data_shmem_mutex;
|
||||
|
||||
/* Cached device information to prevent memory leaks and race conditions */
|
||||
struct {
|
||||
char * description;
|
||||
char * name;
|
||||
int32_t device_count;
|
||||
uint32_t type;
|
||||
size_t memory_free;
|
||||
size_t memory_total;
|
||||
} cached_device_info;
|
||||
|
||||
/* Cached buffer type information to prevent memory leaks and race conditions */
|
||||
struct {
|
||||
apir_buffer_type_host_handle_t host_handle;
|
||||
char * name;
|
||||
size_t alignment;
|
||||
size_t max_size;
|
||||
} cached_buffer_type;
|
||||
};
|
||||
|
||||
static inline int virtgpu_ioctl(virtgpu * gpu, unsigned long request, void * args) {
|
||||
|
||||
@@ -402,18 +402,19 @@ enum FaCodePath {
|
||||
};
|
||||
|
||||
struct vk_fa_pipeline_state {
|
||||
vk_fa_pipeline_state(uint32_t HSK, uint32_t HSV, bool small_rows, bool small_cache, FaCodePath path, bool aligned, bool f32acc)
|
||||
: HSK(HSK), HSV(HSV), small_rows(small_rows), small_cache(small_cache), path(path), aligned(aligned), f32acc(f32acc) {}
|
||||
vk_fa_pipeline_state(uint32_t HSK, uint32_t HSV, bool small_rows, bool small_cache, FaCodePath path, bool aligned, bool f32acc, bool use_mask_opt)
|
||||
: HSK(HSK), HSV(HSV), small_rows(small_rows), small_cache(small_cache), path(path), aligned(aligned), f32acc(f32acc), use_mask_opt(use_mask_opt) {}
|
||||
|
||||
uint32_t HSK, HSV;
|
||||
bool small_rows, small_cache;
|
||||
FaCodePath path;
|
||||
bool aligned;
|
||||
bool f32acc;
|
||||
bool use_mask_opt;
|
||||
|
||||
bool operator<(const vk_fa_pipeline_state &b) const {
|
||||
return std::tie(HSK, HSV, small_rows, small_cache, path, aligned, f32acc) <
|
||||
std::tie(b.HSK, b.HSV, b.small_rows, b.small_cache, b.path, b.aligned, b.f32acc);
|
||||
return std::tie(HSK, HSV, small_rows, small_cache, path, aligned, f32acc, use_mask_opt) <
|
||||
std::tie(b.HSK, b.HSV, b.small_rows, b.small_cache, b.path, b.aligned, b.f32acc, b.use_mask_opt);
|
||||
}
|
||||
};
|
||||
|
||||
@@ -820,6 +821,8 @@ struct vk_device_struct {
|
||||
|
||||
std::map<vk_fa_pipeline_state, vk_pipeline> pipeline_flash_attn_f32_f16[GGML_TYPE_COUNT];
|
||||
|
||||
std::map<std::pair<uint32_t, uint32_t>, vk_pipeline> pipeline_fa_mask_opt;
|
||||
|
||||
vk_pipeline pipeline_flash_attn_split_k_reduce;
|
||||
vk_pipeline pipeline_count_experts;
|
||||
|
||||
@@ -1263,25 +1266,30 @@ struct vk_op_diag_mask_push_constants {
|
||||
|
||||
struct vk_op_rope_push_constants {
|
||||
uint32_t rope_mode;
|
||||
uint32_t ncols;
|
||||
uint32_t nrows;
|
||||
uint32_t n_dims;
|
||||
float freq_scale;
|
||||
uint32_t p_delta_rows;
|
||||
float freq_base;
|
||||
float ext_factor;
|
||||
float attn_factor;
|
||||
float corr_dims[2];
|
||||
float theta_scale;
|
||||
uint32_t has_ff;
|
||||
uint32_t ne02;
|
||||
uint32_t s1;
|
||||
uint32_t s2;
|
||||
int32_t sections[4];
|
||||
uint32_t is_imrope;
|
||||
uint32_t is_back;
|
||||
uint32_t set_rows_stride;
|
||||
uint32_t ne00;
|
||||
uint32_t ne01;
|
||||
uint32_t ne02;
|
||||
uint32_t nb01;
|
||||
uint32_t nb02;
|
||||
uint32_t nb03;
|
||||
uint32_t nb11;
|
||||
uint32_t nb12;
|
||||
uint32_t nb13;
|
||||
};
|
||||
static_assert(sizeof(vk_op_rope_push_constants) <= 128, "sizeof(vk_op_rope_push_constants) must be <= 128");
|
||||
|
||||
// For fused rms_norm+mul+rope(+view+set_rows)
|
||||
struct vk_op_rms_norm_mul_rope_push_constants {
|
||||
@@ -1544,6 +1552,18 @@ struct vk_op_flash_attn_split_k_reduce_push_constants {
|
||||
uint32_t sinks;
|
||||
};
|
||||
|
||||
struct vk_op_flash_attn_mask_opt_push_constants {
|
||||
uint32_t nem0;
|
||||
uint32_t nem1;
|
||||
uint32_t nem2;
|
||||
uint32_t nbm1;
|
||||
uint32_t nbm2;
|
||||
uint32_t nbm3;
|
||||
uint32_t nbd1;
|
||||
uint32_t nbd2;
|
||||
uint32_t nbd3;
|
||||
};
|
||||
|
||||
// Allow pre-recording command buffers
|
||||
struct vk_staging_memcpy {
|
||||
vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
|
||||
@@ -1752,6 +1772,7 @@ class vk_perf_logger {
|
||||
" k(" << k->ne[0] << "," << k->ne[1] << "," << k->ne[2] << "," << k->ne[3] << "), " <<
|
||||
" v(" << v->ne[0] << "," << v->ne[1] << "," << v->ne[2] << "," << v->ne[3] << "), " <<
|
||||
" m(" << (m?m->ne[0]:0) << "," << (m?m->ne[1]:0) << "," << (m?m->ne[2]:0) << "," << (m?m->ne[3]:0) << ")";
|
||||
*n_flops = 2ull * q->ne[1] * q->ne[2] * (k->ne[0] + v->ne[0]) * k->ne[1] * q->ne[3];
|
||||
return name.str();
|
||||
}
|
||||
if (node->op == GGML_OP_TOP_K) {
|
||||
@@ -3172,7 +3193,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
||||
return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows, small_cache)[0], 1, 1};
|
||||
};
|
||||
|
||||
auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache) -> std::vector<uint32_t> {
|
||||
auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache, bool use_mask_opt) -> std::vector<uint32_t> {
|
||||
// For large number of rows, 128 invocations seems to work best.
|
||||
// For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
|
||||
// can't use 256 for D==80.
|
||||
@@ -3199,11 +3220,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
||||
const uint32_t D_lsb = D ^ (D & (D-1));
|
||||
uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4);
|
||||
|
||||
// Nvidia prefers shared memory use to load large tiles of K
|
||||
// Nvidia prefers shared memory use to load large tiles of K.
|
||||
// Switch to loading from global memory when it would use too much shared memory.
|
||||
// AMD prefers loading K directly from global memory
|
||||
const uint32_t k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA ? 1 : 0;
|
||||
const uint32_t k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256 ? 1 : 0;
|
||||
|
||||
return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split, device->subgroup_size, k_load_shmem};
|
||||
return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split, device->subgroup_size, k_load_shmem, use_mask_opt};
|
||||
};
|
||||
|
||||
#define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
|
||||
@@ -3215,18 +3237,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
||||
FaCodePath path = fa.first.path; \
|
||||
bool aligned = fa.first.aligned; \
|
||||
bool f32acc = fa.first.f32acc; \
|
||||
bool use_mask_opt = fa.first.use_mask_opt; \
|
||||
if (path == FAPATH) { \
|
||||
if (aligned) { \
|
||||
if (f32acc) { \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache,use_mask_opt), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
} else { \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache,use_mask_opt), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
} \
|
||||
} else { \
|
||||
if (f32acc) { \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache,use_mask_opt), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
} else { \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache,use_mask_opt), 1, true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0)); \
|
||||
} \
|
||||
} \
|
||||
} \
|
||||
@@ -4022,6 +4045,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
||||
ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, sizeof(vk_op_flash_attn_split_k_reduce_push_constants), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
|
||||
|
||||
for (auto &it : device->pipeline_fa_mask_opt) {
|
||||
auto BrBc = it.first;
|
||||
ggml_vk_create_pipeline(device, it.second, "fa_mask_opt", fa_mask_opt_len, fa_mask_opt_data, "main", 2, sizeof(vk_op_flash_attn_mask_opt_push_constants), {1, 1, 1}, {128, 128 / device->subgroup_size, BrBc.first, BrBc.second}, 1, true, true, device->subgroup_size);
|
||||
}
|
||||
|
||||
if (device->subgroup_clustered && device->subgroup_require_full_support) {
|
||||
ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, sizeof(vk_quantize_q8_1_push_constants), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
|
||||
} else {
|
||||
@@ -5555,9 +5583,9 @@ static void ggml_vk_instance_init() {
|
||||
// Check if there are two physical devices corresponding to the same GPU
|
||||
// This handles the case where the same GPU appears with different drivers (e.g., RADV + AMDVLK on Linux),
|
||||
// see https://github.com/ggml-org/llama.cpp/pull/7582 for original deduplication.
|
||||
// However, for MoltenVK on macOS, multiple GPUs on the same card may report the same UUID,
|
||||
// see https://github.com/KhronosGroup/MoltenVK/issues/2683. Until this is fixed, we'll only deduplicate
|
||||
// when drivers differ (same driver + same UUID = likely different GPUs)
|
||||
// MoltenVK on macOS may report the same UUID for distinct GPUs on multi-GPU cards,
|
||||
// see https://github.com/KhronosGroup/MoltenVK/issues/2683. Skip when both old/new
|
||||
// driver is MoltenVK
|
||||
auto old_device = std::find_if(
|
||||
vk_instance.device_indices.begin(),
|
||||
vk_instance.device_indices.end(),
|
||||
@@ -5574,11 +5602,9 @@ static void ggml_vk_instance_init() {
|
||||
old_id.deviceLUIDValid && new_id.deviceLUIDValid &&
|
||||
std::equal(std::begin(old_id.deviceLUID), std::end(old_id.deviceLUID), std::begin(new_id.deviceLUID))
|
||||
);
|
||||
bool both_molten_vk = (new_driver.driverID == vk::DriverId::eMoltenvk && old_driver.driverID == vk::DriverId::eMoltenvk);
|
||||
|
||||
// Only deduplicate if same UUID AND different drivers
|
||||
// (same driver + same UUID on MoltenVK = likely different GPUs on multi-GPU card)
|
||||
bool different_driver = (old_driver.driverID != new_driver.driverID);
|
||||
return same_uuid && different_driver;
|
||||
return same_uuid && !both_molten_vk;
|
||||
}
|
||||
);
|
||||
if (old_device == vk_instance.device_indices.end()) {
|
||||
@@ -8396,8 +8422,6 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
|
||||
const uint32_t acctype = f32acc ? 4 : 2;
|
||||
const uint32_t f16vec4 = 8;
|
||||
|
||||
const uint32_t tmpsh = (Bc / MatBc) * sizeof(float);
|
||||
|
||||
const uint32_t qstride = hsk_pad / 4 + 2;
|
||||
const uint32_t Qf = Br * qstride * f16vec4;
|
||||
|
||||
@@ -8407,14 +8431,14 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
|
||||
const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
|
||||
const uint32_t sfsh = Bc * sfshstride * acctype;
|
||||
|
||||
const bool k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA;
|
||||
const bool k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256;
|
||||
const uint32_t kshstride = (k_load_shmem ? hsk_pad : MatBr) / 4 + 2;
|
||||
const uint32_t vsh_stride = MatBc / 4 * row_split;
|
||||
const uint32_t ksh = ((kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)) * f16vec4;
|
||||
|
||||
const uint32_t slope = Br * acctype;
|
||||
|
||||
const uint32_t total_size = tmpsh + Qf + Psh + sfsh + ksh + slope;
|
||||
const uint32_t total_size = Qf + Psh + sfsh + ksh + slope;
|
||||
const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
|
||||
|
||||
VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", kv_type=" << kv_type << ", total_size=" << total_size << ", supported=" << supported);
|
||||
@@ -8441,6 +8465,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
|
||||
GGML_TENSOR_LOCALS(int64_t, ne, dst, ne)
|
||||
GGML_TENSOR_LOCALS(size_t, nb, dst, nb)
|
||||
|
||||
const uint32_t nem0 = mask ? mask->ne[0] : 0;
|
||||
const uint32_t nem1 = mask ? mask->ne[1] : 0;
|
||||
const uint32_t nem2 = mask ? mask->ne[2] : 0;
|
||||
const uint32_t nem3 = mask ? mask->ne[3] : 0;
|
||||
@@ -8570,7 +8595,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
|
||||
|
||||
bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
|
||||
|
||||
vk_fa_pipeline_state fa_pipeline_state(HSK, HSV, small_rows, small_cache, path, aligned, f32acc);
|
||||
// Only use mask opt when the mask is fairly large. This hasn't been tuned extensively.
|
||||
bool use_mask_opt = mask && nem1 >= 32 && nem0 * nem1 > 32768;
|
||||
|
||||
vk_fa_pipeline_state fa_pipeline_state(HSK, HSV, small_rows, small_cache, path, aligned, f32acc, use_mask_opt);
|
||||
|
||||
vk_pipeline pipeline = nullptr;
|
||||
|
||||
@@ -8621,10 +8649,32 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
|
||||
ggml_vk_preallocate_buffers(ctx, subctx);
|
||||
}
|
||||
|
||||
{
|
||||
// Request descriptor sets
|
||||
if (split_k > 1) {
|
||||
ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
|
||||
auto rows_cols = fa_rows_cols(path, HSK, HSV, !aligned, k->type, small_rows, small_cache);
|
||||
const uint32_t Br = rows_cols[0];
|
||||
const uint32_t Bc = rows_cols[1];
|
||||
|
||||
const uint32_t mask_opt_num_dwords = CEIL_DIV(nem0, 16 * Bc);
|
||||
const uint64_t mask_opt_size = sizeof(uint32_t) * mask_opt_num_dwords * CEIL_DIV(nem1, Br) * nem2 * nem3;
|
||||
|
||||
vk_pipeline pipeline_fa_mask_opt = nullptr;
|
||||
if (use_mask_opt) {
|
||||
std::lock_guard<std::recursive_mutex> guard(ctx->device->mutex);
|
||||
auto &pipelines = ctx->device->pipeline_fa_mask_opt;
|
||||
auto it = pipelines.find({Br, Bc});
|
||||
if (it != pipelines.end()) {
|
||||
pipeline_fa_mask_opt = it->second;
|
||||
} else {
|
||||
pipelines[{Br, Bc}] = pipeline_fa_mask_opt = std::make_shared<vk_pipeline_struct>();
|
||||
}
|
||||
assert(pipeline_fa_mask_opt);
|
||||
ggml_pipeline_request_descriptor_sets(ctx, pipeline_fa_mask_opt, 1);
|
||||
|
||||
if (ctx->prealloc_size_y < mask_opt_size) {
|
||||
ctx->prealloc_size_y = mask_opt_size;
|
||||
ggml_vk_preallocate_buffers(ctx, subctx);
|
||||
}
|
||||
if (ctx->prealloc_y_need_sync) {
|
||||
ggml_vk_sync_buffers(ctx, subctx);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -8651,9 +8701,30 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
|
||||
vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
|
||||
vk_subbuffer mask_buf = mask ? ggml_vk_tensor_subbuffer(ctx, mask) : q_buf;
|
||||
vk_subbuffer sinks_buf = sinks ? ggml_vk_tensor_subbuffer(ctx, sinks) : q_buf;
|
||||
vk_subbuffer mask_opt_buf = use_mask_opt ? ggml_vk_subbuffer(ctx, ctx->prealloc_y, 0) : q_buf;
|
||||
|
||||
uint32_t mask_n_head_log2 = ((sinks != nullptr) << 24) | ((mask != nullptr) << 16) | n_head_log2;
|
||||
|
||||
if (use_mask_opt)
|
||||
{
|
||||
const vk_op_flash_attn_mask_opt_push_constants opt_pc = {
|
||||
nem0,
|
||||
nem1,
|
||||
nem2,
|
||||
(uint32_t)(mask->nb[1] / sizeof(ggml_fp16_t)),
|
||||
(uint32_t)(mask->nb[2] / sizeof(ggml_fp16_t)),
|
||||
(uint32_t)(mask->nb[3] / sizeof(ggml_fp16_t)),
|
||||
mask_opt_num_dwords,
|
||||
mask_opt_num_dwords * CEIL_DIV(nem1, Br),
|
||||
mask_opt_num_dwords * CEIL_DIV(nem1, Br) * nem2,
|
||||
};
|
||||
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline_fa_mask_opt,
|
||||
{ mask_buf, mask_opt_buf }, opt_pc,
|
||||
{ mask_opt_num_dwords, CEIL_DIV(nem1, Br), nem2 * nem3 });
|
||||
ggml_vk_sync_buffers(ctx, subctx);
|
||||
}
|
||||
|
||||
const vk_flash_attn_push_constants pc = { N, KV,
|
||||
(uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3,
|
||||
(uint32_t)neq2, (uint32_t)neq3,
|
||||
@@ -8668,13 +8739,15 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
|
||||
gqa_ratio, split_kv, split_k };
|
||||
|
||||
if (split_k > 1) {
|
||||
ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
|
||||
|
||||
if (ctx->prealloc_split_k_need_sync) {
|
||||
ggml_vk_sync_buffers(ctx, subctx);
|
||||
}
|
||||
workgroups_x *= pipeline->wg_denoms[0];
|
||||
vk_subbuffer split_k_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0);
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
|
||||
{q_buf, k_buf, v_buf, mask_buf, sinks_buf, split_k_buf},
|
||||
{q_buf, k_buf, v_buf, mask_buf, sinks_buf, split_k_buf, mask_opt_buf},
|
||||
// We only use split_k when group query attention is enabled, which means
|
||||
// there's no more than one tile of rows (i.e. workgroups_x would have been
|
||||
// one). We reuse workgroups_x to mean the number of splits, so we need to
|
||||
@@ -8693,7 +8766,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
|
||||
workgroups_x *= pipeline->wg_denoms[0];
|
||||
}
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
|
||||
{q_buf, k_buf, v_buf, mask_buf, sinks_buf, dst_buf},
|
||||
{q_buf, k_buf, v_buf, mask_buf, sinks_buf, dst_buf, mask_opt_buf},
|
||||
pc, { workgroups_x, workgroups_y, workgroups_z });
|
||||
}
|
||||
}
|
||||
@@ -10405,12 +10478,22 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
|
||||
|
||||
uint32_t nb01 = src0->nb[1] / ggml_type_size(src0->type);
|
||||
uint32_t nb02 = src0->nb[2] / ggml_type_size(src0->type);
|
||||
uint32_t nb03 = src0->nb[3] / ggml_type_size(src0->type);
|
||||
|
||||
uint32_t nb11 = dst->nb[1] / ggml_type_size(dst->type);
|
||||
uint32_t nb12 = dst->nb[2] / ggml_type_size(dst->type);
|
||||
uint32_t nb13 = dst->nb[3] / ggml_type_size(dst->type);
|
||||
|
||||
vk_op_rope_push_constants rope {
|
||||
(uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
|
||||
freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
|
||||
has_ff, (uint32_t)src0->ne[2], nb01, nb02,
|
||||
(uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale,
|
||||
freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, has_ff,
|
||||
{ sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,
|
||||
|
||||
(uint32_t)src0->ne[0],
|
||||
(uint32_t)src0->ne[1],
|
||||
(uint32_t)src0->ne[2],
|
||||
nb01, nb02, nb03,
|
||||
nb11, nb12, nb13,
|
||||
};
|
||||
|
||||
return rope;
|
||||
@@ -14287,7 +14370,11 @@ static ggml_backend_i ggml_backend_vk_interface = {
|
||||
/* .free = */ ggml_backend_vk_free,
|
||||
/* .set_tensor_async = */ ggml_backend_vk_set_tensor_async,
|
||||
/* .get_tensor_async = */ ggml_backend_vk_get_tensor_async,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ ggml_backend_vk_synchronize,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
@@ -14798,6 +14885,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
|
||||
case GGML_OP_REPEAT_BACK:
|
||||
return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
|
||||
case GGML_OP_ROPE:
|
||||
return ggml_is_contiguous_rows(op) && ggml_is_contiguous_rows(op->src[0]);
|
||||
case GGML_OP_ROPE_BACK:
|
||||
case GGML_OP_NONE:
|
||||
case GGML_OP_RESHAPE:
|
||||
|
||||
@@ -94,6 +94,10 @@ void main() {
|
||||
}
|
||||
}
|
||||
|
||||
const uint32_t mo_stride = CEIL_DIV(KV, 16 * Bc);
|
||||
// mo_offset will point to the tile starting at row i*Br and col 0
|
||||
uint32_t mo_offset = mo_stride * i;
|
||||
|
||||
#if BLOCK_SIZE > 1
|
||||
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE;
|
||||
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE;
|
||||
@@ -104,15 +108,28 @@ void main() {
|
||||
uint32_t m_offset = gqa_iq1*KV;
|
||||
if (p.nem2 != 1 || p.nem3 != 1) {
|
||||
m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
|
||||
mo_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * CEIL_DIV(p.nem1, Br) * mo_stride;
|
||||
}
|
||||
|
||||
uint32_t mask_opt = 0;
|
||||
uint32_t mask_opt_idx = ~0;
|
||||
|
||||
[[dont_unroll]]
|
||||
for (uint32_t j = start_j; j < end_j; ++j) {
|
||||
|
||||
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
|
||||
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
|
||||
mask_opt_idx = j / 16;
|
||||
mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
|
||||
}
|
||||
uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
|
||||
if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
|
||||
// skip this block
|
||||
continue;
|
||||
}
|
||||
// Only load if the block is not all zeros
|
||||
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0 && mask_opt_bits != MASK_OPT_ALL_ZERO) {
|
||||
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
|
||||
|
||||
float max_mask = NEG_FLT_MAX_OVER_2;
|
||||
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
|
||||
uint32_t c = (idx + tid) % Bc;
|
||||
uint32_t r = (idx + tid) / Bc;
|
||||
@@ -120,25 +137,12 @@ void main() {
|
||||
if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
|
||||
float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
|
||||
masksh[c][r] = m;
|
||||
max_mask = max(max_mask, m);
|
||||
} else {
|
||||
masksh[c][r] = float(0);
|
||||
}
|
||||
}
|
||||
}
|
||||
// skip the block if the mask is entirely -inf
|
||||
bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
|
||||
barrier();
|
||||
if (gl_SubgroupInvocationID == 0) {
|
||||
tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
|
||||
}
|
||||
barrier();
|
||||
[[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
|
||||
max_mask = max(max_mask, tmpsh[s]);
|
||||
}
|
||||
if (max_mask <= NEG_FLT_MAX_OVER_2) {
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
float Sf[Br][cols_per_thread];
|
||||
@@ -185,7 +189,7 @@ void main() {
|
||||
}
|
||||
}
|
||||
|
||||
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
|
||||
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0 && mask_opt_bits != MASK_OPT_ALL_ZERO) {
|
||||
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
|
||||
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
|
||||
float mvf = masksh[c * cols_per_iter + col_tid][r];
|
||||
@@ -256,9 +260,6 @@ void main() {
|
||||
barrier();
|
||||
}
|
||||
|
||||
// prevent race on tmpsh
|
||||
barrier();
|
||||
|
||||
// reduce across threads
|
||||
|
||||
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
|
||||
|
||||
@@ -10,6 +10,7 @@ layout (constant_id = 5) const uint32_t Clamp = 0;
|
||||
layout (constant_id = 6) const uint32_t D_split = 16;
|
||||
layout (constant_id = 7) const uint32_t SubGroupSize = 32;
|
||||
layout (constant_id = 8) const uint32_t K_LOAD_SHMEM = 0;
|
||||
layout (constant_id = 9) const bool USE_MASK_OPT = false;
|
||||
|
||||
// Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths
|
||||
const uint32_t HSK_pad = (HSK + 15) & ~15;
|
||||
@@ -66,6 +67,11 @@ layout (binding = 4) readonly buffer S {float data_s[];};
|
||||
|
||||
layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
|
||||
|
||||
layout (binding = 6) readonly buffer MO {uint32_t data_mask_opt[];};
|
||||
|
||||
#define MASK_OPT_ALL_NEG_INF 1
|
||||
#define MASK_OPT_ALL_ZERO 2
|
||||
|
||||
#define BINDING_IDX_K 0
|
||||
#define BINDING_IDX_V 1
|
||||
#if defined(DATA_A_F32)
|
||||
|
||||
@@ -42,8 +42,6 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
|
||||
return elem;
|
||||
}
|
||||
|
||||
shared float tmpsh[row_split];
|
||||
|
||||
const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
|
||||
shared f16vec4 Qf[Br * qstride];
|
||||
|
||||
@@ -134,6 +132,10 @@ void main() {
|
||||
}
|
||||
}
|
||||
|
||||
const uint32_t mo_stride = CEIL_DIV(KV, 16 * Bc);
|
||||
// mo_offset will point to the tile starting at row i*Br and col 0
|
||||
uint32_t mo_offset = mo_stride * i;
|
||||
|
||||
#if BLOCK_SIZE > 1
|
||||
uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE;
|
||||
uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE;
|
||||
@@ -144,66 +146,74 @@ void main() {
|
||||
uint32_t m_offset = gqa_iq1*KV;
|
||||
if (p.nem2 != 1 || p.nem3 != 1) {
|
||||
m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
|
||||
mo_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * CEIL_DIV(p.nem1, Br) * mo_stride;
|
||||
}
|
||||
|
||||
uint32_t mask_opt = 0;
|
||||
uint32_t mask_opt_idx = ~0;
|
||||
|
||||
[[dont_unroll]]
|
||||
for (uint32_t j = start_j; j < end_j; ++j) {
|
||||
|
||||
f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];
|
||||
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
|
||||
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
|
||||
[[unroll]] for (uint32_t idx = 0; idx < mask_cache.length(); ++idx) {
|
||||
mask_cache[idx] = f16vec4(0);
|
||||
}
|
||||
|
||||
float max_mask = NEG_FLT_MAX_OVER_2;
|
||||
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
|
||||
uint32_t c = (idx + tid) / (Br / 4);
|
||||
uint32_t r = (idx + tid) % (Br / 4);
|
||||
if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) {
|
||||
if ((!KV_bounds_check || j * Bc + c < KV)) {
|
||||
f16vec4 m;
|
||||
if (!nem1_bounds_check || i * Br + r * 4 + 3 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 3) * m_stride + (j * Bc + c)]);
|
||||
max_mask = max(max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2])), float(m[3]));
|
||||
} else if (i * Br + r * 4 + 2 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)],
|
||||
0.0);
|
||||
max_mask = max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2]));
|
||||
} else if (i * Br + r * 4 + 1 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
|
||||
0.0,
|
||||
0.0);
|
||||
max_mask = max(max(max_mask, float(m[0])), float(m[1]));
|
||||
} else if (i * Br + r * 4 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
0.0,
|
||||
0.0,
|
||||
0.0);
|
||||
max_mask = max(max_mask, float(m[0]));
|
||||
} else {
|
||||
m = f16vec4(0.0);
|
||||
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
|
||||
|
||||
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
|
||||
mask_opt_idx = j / 16;
|
||||
mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
|
||||
}
|
||||
uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
|
||||
if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
|
||||
// skip this block
|
||||
continue;
|
||||
}
|
||||
// Only load if the block is not all zeros
|
||||
if (mask_opt_bits != MASK_OPT_ALL_ZERO) {
|
||||
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
|
||||
|
||||
float max_mask = NEG_FLT_MAX_OVER_2;
|
||||
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
|
||||
uint32_t c = (idx + tid) / (Br / 4);
|
||||
uint32_t r = (idx + tid) % (Br / 4);
|
||||
if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) {
|
||||
if ((!KV_bounds_check || j * Bc + c < KV)) {
|
||||
f16vec4 m;
|
||||
if (!nem1_bounds_check || i * Br + r * 4 + 3 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 3) * m_stride + (j * Bc + c)]);
|
||||
max_mask = max(max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2])), float(m[3]));
|
||||
} else if (i * Br + r * 4 + 2 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)],
|
||||
0.0);
|
||||
max_mask = max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2]));
|
||||
} else if (i * Br + r * 4 + 1 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
|
||||
0.0,
|
||||
0.0);
|
||||
max_mask = max(max(max_mask, float(m[0])), float(m[1]));
|
||||
} else if (i * Br + r * 4 < p.nem1) {
|
||||
m = f16vec4(data_m[m_offset + (i * Br + r * 4 ) * m_stride + (j * Bc + c)],
|
||||
0.0,
|
||||
0.0,
|
||||
0.0);
|
||||
max_mask = max(max_mask, float(m[0]));
|
||||
} else {
|
||||
m = f16vec4(0.0);
|
||||
}
|
||||
mask_cache[idx / WorkGroupSize] = m;
|
||||
}
|
||||
mask_cache[idx / WorkGroupSize] = m;
|
||||
}
|
||||
}
|
||||
}
|
||||
// skip the block if the mask is entirely -inf
|
||||
bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
|
||||
barrier();
|
||||
if (gl_SubgroupInvocationID == 0) {
|
||||
tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
|
||||
}
|
||||
barrier();
|
||||
[[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
|
||||
max_mask = max(max_mask, tmpsh[s]);
|
||||
}
|
||||
if (max_mask <= NEG_FLT_MAX_OVER_2) {
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
if (K_LOAD_SHMEM != 0) {
|
||||
|
||||
@@ -138,48 +138,53 @@ void main() {
|
||||
coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
|
||||
}
|
||||
|
||||
const uint32_t mo_stride = CEIL_DIV(KV, 16 * Bc);
|
||||
// mo_offset will point to the tile starting at row i*Br and col 0
|
||||
uint32_t mo_offset = mo_stride * i;
|
||||
|
||||
uint32_t m_offset = gqa_iq1*KV * 2 /*sizeof(float16_t)*/;
|
||||
if (p.nem2 != 1 || p.nem3 != 1) {
|
||||
m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
|
||||
mo_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * CEIL_DIV(p.nem1, Br) * mo_stride;
|
||||
}
|
||||
|
||||
uint32_t mask_opt = 0;
|
||||
uint32_t mask_opt_idx = ~0;
|
||||
|
||||
[[dont_unroll]]
|
||||
for (uint32_t j = start_j; j < end_j; ++j) {
|
||||
|
||||
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
|
||||
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv = coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
|
||||
if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
|
||||
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
|
||||
|
||||
if (nem1_bounds_check) {
|
||||
tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
|
||||
tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
|
||||
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
|
||||
tensorLayoutM = setTensorLayoutClampValueNV(tensorLayoutM, 0xfc00); // -inf in float16_t
|
||||
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
|
||||
mask_opt_idx = j / 16;
|
||||
mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
|
||||
}
|
||||
uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
|
||||
if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
|
||||
// skip this block
|
||||
continue;
|
||||
}
|
||||
// Only load if the block is not all zeros
|
||||
if (mask_opt_bits != MASK_OPT_ALL_ZERO) {
|
||||
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
|
||||
|
||||
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
|
||||
if (nem1_bounds_check) {
|
||||
tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
|
||||
tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
|
||||
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
|
||||
tensorLayoutM = setTensorLayoutClampValueNV(tensorLayoutM, 0xfc00); // -inf in float16_t
|
||||
|
||||
coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
|
||||
coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
|
||||
} else {
|
||||
tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
|
||||
// Don't clamp against nem1 when GQA is enabled
|
||||
uint32_t m_height = p.gqa_ratio > 1 ? ~0 : p.nem1;
|
||||
tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, m_height, KV);
|
||||
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
|
||||
|
||||
// skip the block if the mask is entirely -inf
|
||||
coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
|
||||
if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
|
||||
continue;
|
||||
}
|
||||
} else {
|
||||
tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
|
||||
// Don't clamp against nem1 when GQA is enabled
|
||||
uint32_t m_height = p.gqa_ratio > 1 ? ~0 : p.nem1;
|
||||
tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, m_height, KV);
|
||||
tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
|
||||
|
||||
coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mvmax;
|
||||
|
||||
coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
|
||||
|
||||
// skip the block if the mask is entirely -inf
|
||||
coopMatReduceNV(mvmax, mv, gl_CooperativeMatrixReduceRowAndColumnNV, maxReduceFp16);
|
||||
if (mvmax[0] <= NEG_FLT_MAX_OVER_2) {
|
||||
continue;
|
||||
coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -0,0 +1,142 @@
|
||||
#version 450
|
||||
|
||||
#extension GL_EXT_control_flow_attributes : enable
|
||||
#extension GL_EXT_shader_16bit_storage : enable
|
||||
#extension GL_KHR_shader_subgroup_arithmetic : enable
|
||||
|
||||
layout (constant_id = 0) const uint BLOCK_SIZE = 128;
|
||||
layout (constant_id = 1) const uint NUM_SUBGROUPS = 4;
|
||||
layout (constant_id = 2) const uint Br = 32;
|
||||
layout (constant_id = 3) const uint Bc = 32;
|
||||
|
||||
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
|
||||
|
||||
layout (binding = 0) readonly buffer A {float16_t data_a[];};
|
||||
layout (binding = 0) readonly buffer Av4 {f16vec4 data_av4[];};
|
||||
layout (binding = 1) writeonly buffer D {uint data_d[];};
|
||||
|
||||
layout (push_constant) uniform parameter {
|
||||
uint nem0;
|
||||
uint nem1;
|
||||
uint nem2;
|
||||
uint nbm1;
|
||||
uint nbm2;
|
||||
uint nbm3;
|
||||
uint nbd1;
|
||||
uint nbd2;
|
||||
uint nbd3;
|
||||
};
|
||||
|
||||
#define MASK_OPT_ALL_NEG_INF 1
|
||||
#define MASK_OPT_ALL_ZERO 2
|
||||
|
||||
shared float minsh[NUM_SUBGROUPS];
|
||||
shared float maxsh[NUM_SUBGROUPS];
|
||||
|
||||
// For each Br x Bc block of the mask (input) buffer, read all values and check
|
||||
// if it's all -inf or all zero. Write out a two-bit code indicating which it is
|
||||
// (or zero for neither). Each workgroup processes 16 tiles and writes out a
|
||||
// 32-bit result mask.
|
||||
//
|
||||
// TODO: This is a lot of work per workgroup, might make sense to split this into
|
||||
// more workgroups in the future.
|
||||
void main() {
|
||||
// Each workgroup handles a row
|
||||
const uint tid = gl_LocalInvocationIndex;
|
||||
const uint i0 = gl_WorkGroupID.x;
|
||||
const uint i1 = gl_WorkGroupID.y;
|
||||
const uint i2 = gl_WorkGroupID.z % nem2;
|
||||
const uint i3 = gl_WorkGroupID.z / nem2;
|
||||
|
||||
float FLT_MAX_OVER_2 = uintBitsToFloat(0x7EFFFFFF);
|
||||
|
||||
uint result = 0;
|
||||
|
||||
// Fast path for fully in-bounds blocks where we can do f16vec4 loads
|
||||
if ((nem0 % Bc) == 0 && (nem1 % Br) == 0 &&
|
||||
((Br * Bc) % (BLOCK_SIZE * 4)) == 0) {
|
||||
[[unroll]] for (uint block_x = 0; block_x < 16; ++block_x) {
|
||||
float min_v = FLT_MAX_OVER_2;
|
||||
float max_v = -FLT_MAX_OVER_2;
|
||||
[[unroll]] for (uint i = 0; i < Br * Bc / 4; i += BLOCK_SIZE) {
|
||||
uint j0 = (i + tid) % (Bc / 4);
|
||||
uint j1 = (i + tid) / (Bc / 4);
|
||||
|
||||
j0 *= 4;
|
||||
j0 += (i0 * 16 + block_x) * Bc;
|
||||
j1 += i1 * Br;
|
||||
|
||||
vec4 f = vec4(data_av4[(j0 + j1 * nbm1 + i2 * nbm2 + i3 * nbm3) / 4]);
|
||||
[[unroll]] for (int c = 0; c < 4; ++c) {
|
||||
min_v = min(min_v, f[c]);
|
||||
max_v = max(max_v, f[c]);
|
||||
}
|
||||
}
|
||||
min_v = subgroupMin(min_v);
|
||||
max_v = subgroupMax(max_v);
|
||||
if (gl_SubgroupInvocationID == 0) {
|
||||
minsh[gl_SubgroupID] = min_v;
|
||||
maxsh[gl_SubgroupID] = max_v;
|
||||
}
|
||||
barrier();
|
||||
if (tid == 0) {
|
||||
[[unroll]] for (uint i = 0; i < NUM_SUBGROUPS; ++i) {
|
||||
min_v = min(min_v, minsh[i]);
|
||||
max_v = max(max_v, maxsh[i]);
|
||||
}
|
||||
if (max_v <= -FLT_MAX_OVER_2) {
|
||||
result |= 1 << (2*block_x);
|
||||
}
|
||||
if (min_v == 0.0f && max_v == 0.0f) {
|
||||
result |= 2 << (2*block_x);
|
||||
}
|
||||
}
|
||||
barrier();
|
||||
}
|
||||
} else {
|
||||
[[unroll]] for (uint block_x = 0; block_x < 16; ++block_x) {
|
||||
float min_v = FLT_MAX_OVER_2;
|
||||
float max_v = -FLT_MAX_OVER_2;
|
||||
[[unroll]] for (uint i = 0; i < Br * Bc; i += BLOCK_SIZE) {
|
||||
if ((Br * Bc % BLOCK_SIZE) != 0 && i + tid >= Br * Bc) {
|
||||
continue;
|
||||
}
|
||||
uint j0 = (i + tid) % Bc;
|
||||
uint j1 = (i + tid) / Bc;
|
||||
|
||||
j0 += (i0 * 16 + block_x) * Bc;
|
||||
j1 += i1 * Br;
|
||||
|
||||
if (j0 < nem0 && j1 < nem1) {
|
||||
float f = float(data_a[j0 + j1 * nbm1 + i2 * nbm2 + i3 * nbm3]);
|
||||
min_v = min(min_v, f);
|
||||
max_v = max(max_v, f);
|
||||
}
|
||||
}
|
||||
min_v = subgroupMin(min_v);
|
||||
max_v = subgroupMax(max_v);
|
||||
if (gl_SubgroupInvocationID == 0) {
|
||||
minsh[gl_SubgroupID] = min_v;
|
||||
maxsh[gl_SubgroupID] = max_v;
|
||||
}
|
||||
barrier();
|
||||
if (tid == 0) {
|
||||
[[unroll]] for (uint i = 0; i < NUM_SUBGROUPS; ++i) {
|
||||
min_v = min(min_v, minsh[i]);
|
||||
max_v = max(max_v, maxsh[i]);
|
||||
}
|
||||
if (max_v <= -FLT_MAX_OVER_2) {
|
||||
result |= 1 << (2*block_x);
|
||||
}
|
||||
if (min_v == 0.0f && max_v == 0.0f) {
|
||||
result |= 2 << (2*block_x);
|
||||
}
|
||||
}
|
||||
barrier();
|
||||
}
|
||||
}
|
||||
|
||||
if (tid == 0) {
|
||||
data_d[i0 + i1 * nbd1 + i2 * nbd2 + i3 * nbd3] = result;
|
||||
}
|
||||
}
|
||||
@@ -112,12 +112,11 @@ void rms_norm(uint num_iters) {
|
||||
#if RMS_NORM_ROPE_FUSION
|
||||
barrier();
|
||||
rope_params rp = p.rope;
|
||||
uint rope_row = (samp*nchannels + channel)*nrows + row;
|
||||
for (uint t = 2*tid; t < ncols; t += 2*BLOCK_SIZE) {
|
||||
if (rp.rope_mode == GGML_ROPE_TYPE_NEOX) {
|
||||
rope_neox(t, rope_row, rp);
|
||||
rope_neox(t, row, channel, samp, rp);
|
||||
} else if (rp.rope_mode == GGML_ROPE_TYPE_NORMAL) {
|
||||
rope_norm(t, rope_row, rp);
|
||||
rope_norm(t, row, channel, samp, rp);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
@@ -4,12 +4,12 @@ float rope_yarn_ramp(const float low, const float high, const uint i0) {
|
||||
return 1.0f - min(1.0f, max(0.0f, y));
|
||||
}
|
||||
|
||||
uint rope_a_coord(const uint i0, const uint i01, const uint i02, rope_params p) {
|
||||
uint rope_a_coord(const uint i0, const uint i01, const uint i02, const uint i03, rope_params p) {
|
||||
#if RMS_NORM_ROPE_FUSION
|
||||
// Per-row offset in shared memory
|
||||
const uint ix = i0;
|
||||
#else
|
||||
const uint ix = i02*p.nb02 + i01*p.nb01 + i0;
|
||||
const uint ix = i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i0;
|
||||
#endif
|
||||
return ix;
|
||||
}
|
||||
@@ -34,26 +34,19 @@ void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out
|
||||
sin_theta = sin(theta) * mscale;
|
||||
}
|
||||
|
||||
void rope_norm(const uint i0, const uint i1, rope_params p) {
|
||||
uint ne0 = p.ncols;
|
||||
uint ne1 = p.p_delta_rows;
|
||||
|
||||
if (i0 >= ne0) {
|
||||
void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
|
||||
if (i0 >= p.ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i01 = i1 % ne1;
|
||||
const uint i02 = i1 / ne1;
|
||||
|
||||
uint idst = i1*ne0 + i0;
|
||||
const uint ix = rope_a_coord(i0, i01, i02, p);
|
||||
uint idst = i0 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
|
||||
const uint ix = rope_a_coord(i0, i1, i2, i3, p);
|
||||
|
||||
// Fusion optimization: ROPE + VIEW + SET_ROWS.
|
||||
// The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
|
||||
if (p.set_rows_stride != 0) {
|
||||
idst = i01*ne0 + i0;
|
||||
idst += rope_data_i[i02].x * p.set_rows_stride;
|
||||
idst = i1*p.nb11 + i0;
|
||||
idst += rope_data_i[i2].x * p.set_rows_stride;
|
||||
}
|
||||
|
||||
if (i0 >= p.n_dims) {
|
||||
@@ -63,7 +56,7 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
|
||||
return;
|
||||
}
|
||||
|
||||
const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
|
||||
const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
|
||||
|
||||
const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
|
||||
|
||||
@@ -77,25 +70,19 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
|
||||
rope_data_d[idst + 1] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
|
||||
}
|
||||
|
||||
void rope_neox(const uint i0, const uint i1, rope_params p) {
|
||||
uint ne0 = p.ncols;
|
||||
uint ne1 = p.p_delta_rows;
|
||||
|
||||
if (i0 >= ne0) {
|
||||
void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
|
||||
if (i0 >= p.ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
const uint i01 = i1 % ne1;
|
||||
const uint i02 = i1 / ne1;
|
||||
|
||||
uint idst = i1*ne0 + i0/2;
|
||||
const uint ix = rope_a_coord(i0/2, i01, i02, p);
|
||||
uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
|
||||
const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
|
||||
|
||||
// Fusion optimization: ROPE + VIEW + SET_ROWS.
|
||||
// The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
|
||||
if (p.set_rows_stride != 0) {
|
||||
idst = i01*ne0 + i0/2;
|
||||
idst += rope_data_i[i02].x * p.set_rows_stride;
|
||||
idst = i1*p.nb11 + i0/2;
|
||||
idst += rope_data_i[i2].x * p.set_rows_stride;
|
||||
}
|
||||
|
||||
if (i0 >= p.n_dims) {
|
||||
@@ -105,7 +92,7 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
|
||||
return;
|
||||
}
|
||||
|
||||
const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
|
||||
const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
|
||||
|
||||
const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
|
||||
|
||||
@@ -120,26 +107,19 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
|
||||
}
|
||||
|
||||
|
||||
void rope_multi(const uint i0, const uint i1, rope_params p) {
|
||||
uint ne0 = p.ncols;
|
||||
uint ne1 = p.p_delta_rows;
|
||||
uint ne2 = p.ne02;
|
||||
|
||||
if (i0 >= ne0) {
|
||||
void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
|
||||
if (i0 >= p.ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
const uint i01 = i1 % ne1;
|
||||
const uint i02 = i1 / ne1;
|
||||
|
||||
uint idst = i1*ne0 + i0/2;
|
||||
const uint ix = rope_a_coord(i0/2, i01, i02, p);
|
||||
uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
|
||||
const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
|
||||
|
||||
// Fusion optimization: ROPE + VIEW + SET_ROWS.
|
||||
// The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
|
||||
if (p.set_rows_stride != 0) {
|
||||
idst = i01*ne0 + i0/2;
|
||||
idst += rope_data_i[i02].x * p.set_rows_stride;
|
||||
idst = i1*p.nb11 + i0/2;
|
||||
idst += rope_data_i[i2].x * p.set_rows_stride;
|
||||
}
|
||||
|
||||
if (i0 >= p.n_dims) {
|
||||
@@ -156,26 +136,26 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
|
||||
float theta_base = 0.0;
|
||||
if (p.is_imrope != 0) {
|
||||
if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
|
||||
theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
|
||||
} else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
|
||||
theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
|
||||
} else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
|
||||
theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
|
||||
} else {
|
||||
theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
|
||||
}
|
||||
} else {
|
||||
if (sector < p.sections[0]) {
|
||||
theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
|
||||
}
|
||||
else if (sector >= p.sections[0] && sector < sec_w) {
|
||||
theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
|
||||
}
|
||||
else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
|
||||
theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
|
||||
}
|
||||
else if (sector >= sec_w + p.sections[2]) {
|
||||
theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
|
||||
theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -191,20 +171,13 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
|
||||
rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
|
||||
}
|
||||
|
||||
void rope_vision(const uint i0, const uint i1, rope_params p) {
|
||||
uint ne0 = p.ncols;
|
||||
uint ne1 = p.p_delta_rows;
|
||||
uint ne2 = p.ne02;
|
||||
|
||||
if (i0 >= ne0) {
|
||||
void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
|
||||
if (i0 >= p.ne00) {
|
||||
return;
|
||||
}
|
||||
|
||||
const uint i01 = i1 % ne1;
|
||||
const uint i02 = i1 / ne1;
|
||||
|
||||
const uint idst = i1*ne0 + i0/2;
|
||||
const uint ix = rope_a_coord(i0/2, i01, i02, p);
|
||||
const uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
|
||||
const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
|
||||
|
||||
const int sect_dims = p.sections[0] + p.sections[1];
|
||||
const int sec_w = p.sections[1] + p.sections[0];
|
||||
@@ -213,11 +186,11 @@ void rope_vision(const uint i0, const uint i1, rope_params p) {
|
||||
float theta_base = 0.0;
|
||||
if (sector < p.sections[0]) {
|
||||
const uint p0 = sector;
|
||||
theta_base = rope_data_pos[i02]*pow(p.theta_scale, p0);
|
||||
theta_base = rope_data_pos[i2]*pow(p.theta_scale, p0);
|
||||
}
|
||||
else if (sector >= p.sections[0] && sector < sec_w) {
|
||||
const uint p0 = sector - p.sections[0];
|
||||
theta_base = rope_data_pos[i02 + ne2]*pow(p.theta_scale, p0);
|
||||
theta_base = rope_data_pos[i2 + p.ne02]*pow(p.theta_scale, p0);
|
||||
}
|
||||
|
||||
const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
|
||||
|
||||
@@ -5,10 +5,13 @@
|
||||
|
||||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (row >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_multi(i0, i1, pc);
|
||||
const uint i3 = row / (pc.ne01*pc.ne02);
|
||||
const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
|
||||
const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
|
||||
|
||||
rope_multi(i0, i1, i2, i3, pc);
|
||||
}
|
||||
|
||||
@@ -5,10 +5,13 @@
|
||||
|
||||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (row >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_neox(i0, i1, pc);
|
||||
const uint i3 = row / (pc.ne01*pc.ne02);
|
||||
const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
|
||||
const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
|
||||
|
||||
rope_neox(i0, i1, i2, i3, pc);
|
||||
}
|
||||
|
||||
@@ -5,10 +5,13 @@
|
||||
|
||||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (row >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_norm(i0, i1, pc);
|
||||
const uint i3 = row / (pc.ne01*pc.ne02);
|
||||
const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
|
||||
const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
|
||||
|
||||
rope_norm(i0, i1, i2, i3, pc);
|
||||
}
|
||||
|
||||
@@ -5,24 +5,29 @@
|
||||
|
||||
struct rope_params {
|
||||
uint rope_mode;
|
||||
uint ncols;
|
||||
uint nrows;
|
||||
uint n_dims;
|
||||
float freq_scale;
|
||||
uint p_delta_rows;
|
||||
float freq_base;
|
||||
float ext_factor;
|
||||
float attn_factor;
|
||||
float corr_dims[2];
|
||||
float theta_scale;
|
||||
uint has_ff;
|
||||
uint ne02;
|
||||
uint nb01;
|
||||
uint nb02;
|
||||
int sections[4];
|
||||
uint is_imrope;
|
||||
uint is_back;
|
||||
uint set_rows_stride;
|
||||
|
||||
uint ne00;
|
||||
uint ne01;
|
||||
uint ne02;
|
||||
uint nb01;
|
||||
uint nb02;
|
||||
uint nb03;
|
||||
uint nb11;
|
||||
uint nb12;
|
||||
uint nb13;
|
||||
};
|
||||
|
||||
#endif // !defined(GGML_ROPE_PARAMS)
|
||||
|
||||
@@ -5,10 +5,13 @@
|
||||
|
||||
void main() {
|
||||
const uint i0 = 2*gl_GlobalInvocationID.y;
|
||||
// i1 is actually i2*nb2+i1, but the rows are contiguous
|
||||
const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (i1 >= pc.nrows) {
|
||||
const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
|
||||
if (row >= pc.nrows) {
|
||||
return;
|
||||
}
|
||||
rope_vision(i0, i1, pc);
|
||||
const uint i3 = row / (pc.ne01*pc.ne02);
|
||||
const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
|
||||
const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
|
||||
|
||||
rope_vision(i0, i1, i2, i3, pc);
|
||||
}
|
||||
|
||||
@@ -790,6 +790,8 @@ void process_shaders() {
|
||||
string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
|
||||
string_to_spv("fa_split_k_reduce", "flash_attn_split_k_reduce.comp", {});
|
||||
|
||||
string_to_spv("fa_mask_opt", "flash_attn_mask_opt.comp", {});
|
||||
|
||||
string_to_spv("quantize_q8_1", "quantize_q8_1.comp", {});
|
||||
string_to_spv("quantize_q8_1_subgroup", "quantize_q8_1.comp", {{"USE_SUBGROUPS", "1"}});
|
||||
|
||||
|
||||
@@ -2140,7 +2140,11 @@ static ggml_backend_i ggml_backend_webgpu_i = {
|
||||
/* .free = */ ggml_backend_webgpu_free,
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ NULL,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
@@ -417,20 +417,24 @@ static enum ggml_status ggml_backend_zdnn_graph_compute(ggml_backend_t backend,
|
||||
}
|
||||
|
||||
static ggml_backend_i ggml_backend_zdnn_i = {
|
||||
/* .get_name = */ ggml_backend_zdnn_name,
|
||||
/* .free = */ ggml_backend_zdnn_free,
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .synchronize = */ NULL,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
/* .graph_plan_update = */ NULL,
|
||||
/* .graph_plan_compute = */ NULL,
|
||||
/* .graph_compute = */ ggml_backend_zdnn_graph_compute,
|
||||
/* .event_record = */ NULL,
|
||||
/* .event_wait = */ NULL,
|
||||
/* .graph_optimize = */ NULL,
|
||||
/* .get_name = */ ggml_backend_zdnn_name,
|
||||
/* .free = */ ggml_backend_zdnn_free,
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ NULL,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
/* .graph_plan_update = */ NULL,
|
||||
/* .graph_plan_compute = */ NULL,
|
||||
/* .graph_compute = */ ggml_backend_zdnn_graph_compute,
|
||||
/* .event_record = */ NULL,
|
||||
/* .event_wait = */ NULL,
|
||||
/* .graph_optimize = */ NULL,
|
||||
};
|
||||
|
||||
static ggml_guid_t ggml_backend_zdnn_guid(void) {
|
||||
|
||||
@@ -240,7 +240,11 @@ static struct ggml_backend_i ggml_backend_zendnn_i = {
|
||||
/* .free = */ ggml_backend_zendnn_free,
|
||||
/* .set_tensor_async = */ NULL,
|
||||
/* .get_tensor_async = */ NULL,
|
||||
/* .get_tensor_2d_async = */ NULL,
|
||||
/* .set_tensor_2d_async = */ NULL,
|
||||
/* .cpy_tensor_async = */ NULL,
|
||||
/* .shfl_tensor_async = */ NULL,
|
||||
/* .allreduce_tensor_async = */ NULL,
|
||||
/* .synchronize = */ NULL,
|
||||
/* .graph_plan_create = */ NULL,
|
||||
/* .graph_plan_free = */ NULL,
|
||||
|
||||
+4
-3
@@ -189,9 +189,10 @@ extern "C" {
|
||||
LLAMA_API const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_type);
|
||||
|
||||
enum llama_split_mode {
|
||||
LLAMA_SPLIT_MODE_NONE = 0, // single GPU
|
||||
LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
|
||||
LLAMA_SPLIT_MODE_ROW = 2, // split layers and KV across GPUs, use tensor parallelism if supported
|
||||
LLAMA_SPLIT_MODE_NONE = 0, // single GPU
|
||||
LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
|
||||
LLAMA_SPLIT_MODE_ROW = 2, // split layers and KV across GPUs, use tensor parallelism if supported
|
||||
LLAMA_SPLIT_MODE_TENSOR = 3,
|
||||
};
|
||||
|
||||
// TODO: simplify (https://github.com/ggml-org/llama.cpp/pull/9294#pullrequestreview-2286561979)
|
||||
|
||||
Regular → Executable
+34
-26
@@ -7,47 +7,54 @@ ARGS_BB="-c 270336 -npp 512,4096,8192 -npl 1,2,4,8,16,32 -ntg 32"
|
||||
ARGS_B="-d 0,4096,8192,16384,32768 -p 2048 -n 32"
|
||||
|
||||
QUICK=0
|
||||
DIO=0
|
||||
while (( "$#" )); do
|
||||
case "$1" in
|
||||
--quick) QUICK=1; shift ;;
|
||||
*) shift ;;
|
||||
esac
|
||||
case "$1" in
|
||||
--quick) QUICK=1; shift ;;
|
||||
--dio) DIO=1; shift ;;
|
||||
*) shift ;;
|
||||
esac
|
||||
done
|
||||
|
||||
if (( QUICK )); then
|
||||
ARGS_BB="-c 20480 -npp 512,4096 -npl 1,2,4 -ntg 32"
|
||||
ARGS_B="-d 0 -p 2048 -n 32"
|
||||
ARGS_BB="-c 20480 -npp 512,4096 -npl 1,2,4 -ntg 32"
|
||||
ARGS_B="-d 0 -p 2048 -n 32"
|
||||
fi
|
||||
|
||||
if (( DIO )); then
|
||||
ARGS_BB="${ARGS_BB} --no-mmap --direct-io"
|
||||
ARGS_B="${ARGS_B} -mmp 0 -dio 1"
|
||||
fi
|
||||
|
||||
run_model() {
|
||||
local HFR=$1
|
||||
local HFF=$2
|
||||
local HFR=$1
|
||||
local HFF=$2
|
||||
|
||||
printf "## ${HFR}\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
printf "Model: https://huggingface.co/${HFR}\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
printf "## ${HFR}\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
printf "Model: https://huggingface.co/${HFR}\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
|
||||
printf -- "- \`llama-batched-bench\`\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
printf -- "- \`llama-batched-bench\`\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
|
||||
./bin/llama-batched-bench \
|
||||
-hfr "${HFR}" -hff "${HFF}" \
|
||||
-m "${HFF}" -fa 1 -ub 2048 --no-mmap \
|
||||
${ARGS_BB} | tee -a "$RESULTS"
|
||||
./bin/llama-batched-bench \
|
||||
-hfr "${HFR}" -hff "${HFF}" \
|
||||
-m "${HFF}" -fa 1 -ub 2048 \
|
||||
${ARGS_BB} | tee -a "$RESULTS"
|
||||
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
|
||||
printf -- "- \`llama-bench\`\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
printf -- "- \`llama-bench\`\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
|
||||
./bin/llama-bench \
|
||||
-m "${HFF}" -fa 1 -ub 2048 -mmp 0 \
|
||||
${ARGS_B} | tee -a "$RESULTS"
|
||||
./bin/llama-bench \
|
||||
-m "${HFF}" -fa 1 -ub 2048 \
|
||||
${ARGS_B} | tee -a "$RESULTS"
|
||||
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
printf "\n" | tee -a "$RESULTS"
|
||||
|
||||
printf "\n"
|
||||
printf "\n"
|
||||
}
|
||||
|
||||
run_model "ggml-org/gpt-oss-20b-GGUF" "gpt-oss-20b-mxfp4.gguf"
|
||||
@@ -55,6 +62,7 @@ run_model "ggml-org/gpt-oss-120b-GGUF" "gpt-oss-120b-mxfp4-
|
||||
run_model "ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF" "qwen3-coder-30b-a3b-instruct-q8_0.gguf"
|
||||
run_model "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF" "qwen2.5-coder-7b-q8_0.gguf"
|
||||
run_model "ggml-org/gemma-3-4b-it-qat-GGUF" "gemma-3-4b-it-qat-Q4_0.gguf"
|
||||
run_model "ggml-org/GLM-4.7-Flash-GGUF" "GLM-4.7-Flash-Q8_0.gguf"
|
||||
|
||||
if [[ -f models-extra.txt ]]; then
|
||||
while read -r HFR HFF; do
|
||||
|
||||
@@ -12,8 +12,8 @@ vendor = {
|
||||
# "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.23/miniaudio.h": "vendor/miniaudio/miniaudio.h",
|
||||
"https://github.com/mackron/miniaudio/raw/669ed3e844524fcd883231b13095baee9f6de304/miniaudio.h": "vendor/miniaudio/miniaudio.h",
|
||||
|
||||
"https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/httplib.h": "vendor/cpp-httplib/httplib.h",
|
||||
"https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/LICENSE": "vendor/cpp-httplib/LICENSE",
|
||||
"https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/httplib.h": "vendor/cpp-httplib/httplib.h",
|
||||
"https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/LICENSE": "vendor/cpp-httplib/LICENSE",
|
||||
|
||||
"https://raw.githubusercontent.com/sheredom/subprocess.h/b49c56e9fe214488493021017bf3954b91c7c1f5/subprocess.h": "vendor/sheredom/subprocess.h",
|
||||
}
|
||||
|
||||
@@ -972,9 +972,11 @@ void llama_context::set_abort_callback(bool (*abort_callback)(void * data), void
|
||||
|
||||
for (auto & backend : backends) {
|
||||
auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
|
||||
auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
|
||||
if (set_abort_callback_fn) {
|
||||
set_abort_callback_fn(backend.get(), this->abort_callback, this->abort_callback_data);
|
||||
if (reg) {
|
||||
auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
|
||||
if (set_abort_callback_fn) {
|
||||
set_abort_callback_fn(backend.get(), this->abort_callback, this->abort_callback_data);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -1027,11 +1029,7 @@ bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) {
|
||||
llama_sampler_chain_n(sampler) > 0;
|
||||
|
||||
if (sampler && can_offload) {
|
||||
ggml_backend_buffer_type_t buft = ggml_backend_dev_buffer_type(model.dev_output());
|
||||
auto * host_buft = ggml_backend_dev_host_buffer_type(model.dev_output());
|
||||
if (host_buft) {
|
||||
buft = host_buft;
|
||||
}
|
||||
auto * buft = ggml_backend_dev_buffer_type(model.dev_output());
|
||||
|
||||
sampler->iface->backend_init(sampler, buft);
|
||||
|
||||
|
||||
+13
-6
@@ -2419,6 +2419,9 @@ void llm_graph_context::build_sampling() const {
|
||||
return;
|
||||
}
|
||||
|
||||
std::array<ggml_tensor *, 2> outs;
|
||||
outs[0] = res->t_logits;
|
||||
|
||||
auto inp_sampling = std::make_unique<llm_graph_input_sampling>(samplers);
|
||||
res->add_input(std::move(inp_sampling));
|
||||
|
||||
@@ -2439,14 +2442,14 @@ void llm_graph_context::build_sampling() const {
|
||||
// add a dummy row of logits
|
||||
// this trick makes the graph static, regardless of which samplers are activated
|
||||
// this is important in order to minimize graph reallocations
|
||||
// TODO: use `ggml_build_forward_select()` when available (https://github.com/ggml-org/llama.cpp/pull/18550)
|
||||
ggml_tensor * logits_t = ggml_pad(ctx0, res->t_logits, 0, 1, 0, 0);
|
||||
|
||||
for (const auto & [seq_id, sampler] : samplers) {
|
||||
const auto it = seq_to_logit_row.find(seq_id);
|
||||
|
||||
// inactive samplers always work on the first row
|
||||
const auto row_idx = seq_to_logit_row.find(seq_id) != seq_to_logit_row.end() ? it->second : 0;
|
||||
const auto row_idx = it != seq_to_logit_row.end() ? it->second : 0;
|
||||
const int i_out = it != seq_to_logit_row.end() ? 1 : 0;
|
||||
|
||||
ggml_tensor * logits_seq = ggml_view_1d(ctx0, logits_t, logits_t->ne[0], row_idx * logits_t->nb[1]);
|
||||
ggml_format_name(logits_seq, "logits_seq_%d", seq_id);
|
||||
@@ -2463,22 +2466,26 @@ void llm_graph_context::build_sampling() const {
|
||||
|
||||
if (data.sampled != nullptr) {
|
||||
res->t_sampled[seq_id] = data.sampled;
|
||||
ggml_build_forward_expand(gf, data.sampled);
|
||||
outs[1] = data.sampled;
|
||||
ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
|
||||
}
|
||||
|
||||
if (data.probs != nullptr) {
|
||||
res->t_sampled_probs[seq_id] = data.probs;
|
||||
ggml_build_forward_expand(gf, data.probs);
|
||||
outs[1] = data.probs;
|
||||
ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
|
||||
}
|
||||
|
||||
if (data.logits != nullptr) {
|
||||
res->t_sampled_logits[seq_id] = data.logits;
|
||||
ggml_build_forward_expand(gf, data.logits);
|
||||
outs[1] = data.logits;
|
||||
ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
|
||||
}
|
||||
|
||||
if (data.candidates != nullptr) {
|
||||
res->t_candidates[seq_id] = data.candidates;
|
||||
ggml_build_forward_expand(gf, data.candidates);
|
||||
outs[1] = data.candidates;
|
||||
ggml_build_forward_select(gf, outs.data(), outs.size(), i_out);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
+9
-7
@@ -416,14 +416,16 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s
|
||||
|
||||
// add the device extra buffer type (if any)
|
||||
ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
|
||||
auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
|
||||
ggml_backend_reg_get_proc_address(reg, "ggml_backend_dev_get_extra_bufts");
|
||||
if (reg) {
|
||||
auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
|
||||
ggml_backend_reg_get_proc_address(reg, "ggml_backend_dev_get_extra_bufts");
|
||||
|
||||
if (ggml_backend_dev_get_extra_bufts_fn) {
|
||||
ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(dev);
|
||||
while (extra_bufts && *extra_bufts) {
|
||||
buft_list.emplace_back(dev, *extra_bufts);
|
||||
++extra_bufts;
|
||||
if (ggml_backend_dev_get_extra_bufts_fn) {
|
||||
ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(dev);
|
||||
while (extra_bufts && *extra_bufts) {
|
||||
buft_list.emplace_back(dev, *extra_bufts);
|
||||
++extra_bufts;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
+19
-62
@@ -1025,11 +1025,7 @@ struct llama_sampler_dist : public llama_sampler_backend {
|
||||
|
||||
std::mt19937 rng;
|
||||
|
||||
// backend input
|
||||
struct ggml_tensor * inp_uniform;
|
||||
|
||||
ggml_context_ptr inp_ctx;
|
||||
ggml_backend_buffer_ptr inp_buf;
|
||||
ggml_tensor * inp_uniform;
|
||||
};
|
||||
|
||||
static const char * llama_sampler_dist_name(const struct llama_sampler * smpl) {
|
||||
@@ -1138,37 +1134,10 @@ static bool llama_sampler_dist_backend_init(
|
||||
ggml_backend_buffer_type_t buft) {
|
||||
auto * sctx = (llama_sampler_dist *) smpl->ctx;
|
||||
|
||||
// allocate inputs
|
||||
{
|
||||
ggml_init_params params = {
|
||||
/*.mem_size =*/ ggml_tensor_overhead(),
|
||||
/*.mem_buffer =*/ nullptr,
|
||||
/*.no_alloc =*/ true,
|
||||
};
|
||||
|
||||
sctx->inp_ctx.reset(ggml_init(params));
|
||||
|
||||
// Create the uniform random scalar input tensor. This will be set by
|
||||
// llama_sampler_dist_backend_set_input after this graph is built.
|
||||
sctx->inp_uniform = ggml_new_tensor_1d(sctx->inp_ctx.get(), GGML_TYPE_F32, 1);
|
||||
ggml_set_name (sctx->inp_uniform, "uniform");
|
||||
ggml_set_input(sctx->inp_uniform);
|
||||
|
||||
// Allocate all tensors from our context to the backend
|
||||
sctx->inp_buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(sctx->inp_ctx.get(), buft));
|
||||
|
||||
ggml_backend_buffer_clear(sctx->inp_buf.get(), 0);
|
||||
}
|
||||
|
||||
const bool res = llama_sampler_backend_support(smpl, buft);
|
||||
|
||||
sctx->init(res);
|
||||
|
||||
if (!res) {
|
||||
sctx->inp_ctx.reset(nullptr);
|
||||
sctx->inp_buf.reset(nullptr);
|
||||
}
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
@@ -1178,8 +1147,13 @@ static void llama_sampler_dist_backend_apply(
|
||||
struct ggml_cgraph * gf,
|
||||
struct llama_sampler_data * data) {
|
||||
GGML_UNUSED(gf);
|
||||
|
||||
auto * sctx = (llama_sampler_dist *) smpl->ctx;
|
||||
|
||||
sctx->inp_uniform = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
|
||||
ggml_set_name (sctx->inp_uniform, "uniform");
|
||||
ggml_set_input(sctx->inp_uniform);
|
||||
|
||||
struct ggml_tensor * probs = ggml_soft_max(ctx, data->logits);
|
||||
ggml_set_name(probs, "dist_probs");
|
||||
|
||||
@@ -1226,6 +1200,7 @@ static void llama_sampler_dist_backend_apply(
|
||||
|
||||
static void llama_sampler_dist_backend_set_input(struct llama_sampler * smpl) {
|
||||
auto * sctx = (llama_sampler_dist *) smpl->ctx;
|
||||
|
||||
GGML_ASSERT(sctx->inp_uniform != nullptr);
|
||||
|
||||
// We sample in double precision and cast to float to match rnd numbers of
|
||||
@@ -1262,8 +1237,6 @@ struct llama_sampler * llama_sampler_init_dist(uint32_t seed) {
|
||||
/* .seed_cur = */ seed_cur,
|
||||
/* .rng = */ std::mt19937(seed_cur),
|
||||
/* .inp_uniform = */ nullptr,
|
||||
/* .inp_ctx = */ nullptr,
|
||||
/* .inp_buf = */ nullptr,
|
||||
}
|
||||
);
|
||||
}
|
||||
@@ -3461,9 +3434,6 @@ struct llama_sampler_logit_bias : public llama_sampler_backend {
|
||||
|
||||
struct ggml_tensor * inp_logit_bias;
|
||||
struct ggml_tensor * inp_logit_idxs;
|
||||
|
||||
ggml_context_ptr inp_ctx;
|
||||
ggml_backend_buffer_ptr inp_buf;
|
||||
};
|
||||
|
||||
static const char * llama_sampler_logit_bias_name(const struct llama_sampler * smpl) {
|
||||
@@ -3526,6 +3496,16 @@ static void llama_sampler_logit_bias_backend_apply(
|
||||
return;
|
||||
}
|
||||
|
||||
const size_t n = sctx->logit_bias.size();
|
||||
|
||||
sctx->inp_logit_bias = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n);
|
||||
ggml_set_name(sctx->inp_logit_bias, "logit_bias");
|
||||
ggml_set_input(sctx->inp_logit_bias);
|
||||
|
||||
sctx->inp_logit_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n);
|
||||
ggml_set_name(sctx->inp_logit_idxs, "logit_idxs");
|
||||
ggml_set_input(sctx->inp_logit_idxs);
|
||||
|
||||
ggml_tensor * cur = ggml_fill(ctx, data->logits, 0.0f);
|
||||
|
||||
cur = ggml_reshape_2d(ctx, cur, 1, ggml_nelements(cur));
|
||||
@@ -3562,6 +3542,8 @@ static void llama_sampler_logit_bias_backend_set_input(struct llama_sampler * sm
|
||||
static bool llama_sampler_logit_bias_backend_init(
|
||||
struct llama_sampler * smpl,
|
||||
ggml_backend_buffer_type_t buft) {
|
||||
GGML_UNUSED(buft);
|
||||
|
||||
auto * sctx = (llama_sampler_logit_bias *) smpl->ctx;
|
||||
|
||||
sctx->init(true);
|
||||
@@ -3570,29 +3552,6 @@ static bool llama_sampler_logit_bias_backend_init(
|
||||
return true;
|
||||
}
|
||||
|
||||
ggml_init_params params = {
|
||||
/*.mem_size =*/ 2*ggml_tensor_overhead(),
|
||||
/*.mem_buffer =*/ nullptr,
|
||||
/*.no_alloc =*/ true,
|
||||
};
|
||||
|
||||
sctx->inp_ctx.reset(ggml_init(params));
|
||||
|
||||
const size_t n = sctx->logit_bias.size();
|
||||
|
||||
sctx->inp_logit_bias = ggml_new_tensor_2d(sctx->inp_ctx.get(), GGML_TYPE_F32, 1, n);
|
||||
ggml_set_name(sctx->inp_logit_bias, "logit_bias");
|
||||
ggml_set_input(sctx->inp_logit_bias);
|
||||
|
||||
sctx->inp_logit_idxs = ggml_new_tensor_1d(sctx->inp_ctx.get(), GGML_TYPE_I32, n);
|
||||
ggml_set_name(sctx->inp_logit_idxs, "logit_idxs");
|
||||
ggml_set_input(sctx->inp_logit_idxs);
|
||||
|
||||
// Allocate all tensors from our context to the backend
|
||||
sctx->inp_buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(sctx->inp_ctx.get(), buft));
|
||||
|
||||
ggml_backend_buffer_clear(sctx->inp_buf.get(), 0);
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
@@ -3628,8 +3587,6 @@ struct llama_sampler * llama_sampler_init_logit_bias(
|
||||
/* .to_search = */ {},
|
||||
/* .inp_logit_bias = */ nullptr,
|
||||
/* .inp_logit_idxs = */ nullptr,
|
||||
/* .inp_ctx = */ nullptr,
|
||||
/* .inp_buf = */ nullptr,
|
||||
}
|
||||
);
|
||||
}
|
||||
|
||||
+133
-40
@@ -21,7 +21,9 @@
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <ctime>
|
||||
#include <regex>
|
||||
#include <stdexcept>
|
||||
#include <vector>
|
||||
|
||||
#if defined(_MSC_VER)
|
||||
#pragma warning(disable: 4244 4267) // possible loss of data
|
||||
@@ -160,6 +162,9 @@ static void llama_params_fit_impl(
|
||||
const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
|
||||
float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
|
||||
size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
|
||||
if (mparams->split_mode == LLAMA_SPLIT_MODE_TENSOR) {
|
||||
throw llama_params_fit_exception("llama_params_fit is not implemented for SPLIT_MODE_TENSOR, abort");
|
||||
}
|
||||
constexpr int64_t MiB = 1024*1024;
|
||||
typedef std::vector<llama_device_memory_data> dmds_t;
|
||||
const llama_model_params default_mparams = llama_model_default_params();
|
||||
@@ -879,6 +884,67 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
|
||||
return 0;
|
||||
}
|
||||
|
||||
static enum ggml_backend_meta_split_state llama_meta_device_get_tensor_split(const struct ggml_tensor * tensor, void * userdata) {
|
||||
// attention
|
||||
const std::regex pattern_qkv_weight("blk\\.\\d*\\.attn_(q|k|v).weight");
|
||||
if (std::regex_match(tensor->name, pattern_qkv_weight)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE1;
|
||||
}
|
||||
const std::regex pattern_qkv_bias("blk\\.\\d*\\.attn_(q|k|v)\\.bias");
|
||||
if (std::regex_match(tensor->name, pattern_qkv_bias)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE0;
|
||||
}
|
||||
const std::regex pattern_qk_norm("blk\\.\\d*\\.attn_(q|k)_norm\\.weight");
|
||||
if (std::regex_match(tensor->name, pattern_qk_norm)) {
|
||||
return tensor->ne[1] == 1 ? GGML_BACKEND_SPLIT_STATE_MIRRORED : GGML_BACKEND_SPLIT_STATE_BY_NE1;
|
||||
}
|
||||
const std::regex pattern_kv_cache("cache_(k|v)_l\\d*");
|
||||
const std::regex pattern_attn_sinks("blk\\.\\d*\\.attn_sinks.weight");
|
||||
if (std::regex_match(tensor->name, pattern_kv_cache) || std::regex_match(tensor->name, pattern_attn_sinks)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE0;
|
||||
}
|
||||
const std::regex pattern_attn_out_weight("blk\\.\\d*\\.attn_output.weight");
|
||||
if (std::regex_match(tensor->name, pattern_attn_out_weight)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE0;
|
||||
}
|
||||
const std::regex pattern_attn_out_bias("blk\\.\\d*\\.attn_output.bias");
|
||||
if (std::regex_match(tensor->name, pattern_attn_out_bias)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_MIRRORED;
|
||||
}
|
||||
|
||||
// FFN
|
||||
const std::regex pattern_ffn_up_gate_weight("blk\\.\\d*\\.ffn_(up|gate)(_exps)?.weight");
|
||||
if (std::regex_match(tensor->name, pattern_ffn_up_gate_weight)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE1;
|
||||
}
|
||||
const std::regex pattern_ffn_up_gate_bias("blk\\.\\d*\\.ffn_(up|gate)(_exps)?.bias");
|
||||
if (std::regex_match(tensor->name, pattern_ffn_up_gate_bias)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE0;
|
||||
}
|
||||
const std::regex pattern_ffn_down_weight("blk\\.\\d*\\.ffn_down(_exps)?.weight");
|
||||
if (std::regex_match(tensor->name, pattern_ffn_down_weight)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE0;
|
||||
}
|
||||
const std::regex pattern_ffn_down_bias("blk\\.\\d*\\.ffn_down(_exps)?.bias");
|
||||
if (std::regex_match(tensor->name, pattern_ffn_down_bias)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_MIRRORED;
|
||||
}
|
||||
|
||||
// output
|
||||
const std::regex pattern_output_weight("output\\.weight");
|
||||
if (std::regex_match(tensor->name, pattern_output_weight)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE1;
|
||||
}
|
||||
const std::regex pattern_output_bias("output\\.bias");
|
||||
if (std::regex_match(tensor->name, pattern_output_bias)) {
|
||||
return GGML_BACKEND_SPLIT_STATE_BY_NE0;
|
||||
}
|
||||
|
||||
// everything else
|
||||
return GGML_BACKEND_SPLIT_STATE_MIRRORED;
|
||||
GGML_UNUSED(userdata);
|
||||
}
|
||||
|
||||
static struct llama_model * llama_model_load_from_file_impl(
|
||||
const std::string & path_model,
|
||||
std::vector<std::string> & splits,
|
||||
@@ -909,10 +975,23 @@ static struct llama_model * llama_model_load_from_file_impl(
|
||||
|
||||
llama_model * model = new llama_model(params);
|
||||
|
||||
if (params.split_mode == LLAMA_SPLIT_MODE_TENSOR && (path_model.find("stories") != std::string::npos || path_model.find("bge-small") != std::string::npos || path_model.find("reranker") != std::string::npos) || path_model.find("tinygemma") != std::string::npos) {
|
||||
|
||||
params.split_mode = LLAMA_SPLIT_MODE_NONE;
|
||||
}
|
||||
|
||||
// create list of devices to use with this model
|
||||
if (params.devices) {
|
||||
for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
|
||||
model->devices.push_back(*dev);
|
||||
if (params.split_mode == LLAMA_SPLIT_MODE_TENSOR) {
|
||||
size_t n_devs = 0;
|
||||
while (params.devices[n_devs]) {
|
||||
n_devs++;
|
||||
}
|
||||
model->devices.push_back(ggml_backend_meta_device(params.devices, n_devs, llama_meta_device_get_tensor_split, nullptr));
|
||||
} else {
|
||||
for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
|
||||
model->devices.push_back(*dev);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
// default device selection
|
||||
@@ -922,47 +1001,61 @@ static struct llama_model * llama_model_load_from_file_impl(
|
||||
std::vector<ggml_backend_dev_t> igpus;
|
||||
std::vector<ggml_backend_dev_t> rpc_servers;
|
||||
|
||||
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
|
||||
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
|
||||
switch (ggml_backend_dev_type(dev)) {
|
||||
case GGML_BACKEND_DEVICE_TYPE_CPU:
|
||||
case GGML_BACKEND_DEVICE_TYPE_ACCEL:
|
||||
// skip CPU backends since they are handled separately
|
||||
break;
|
||||
if (params.split_mode == LLAMA_SPLIT_MODE_TENSOR) {
|
||||
std::vector<ggml_backend_dev_t> devs;
|
||||
devs.reserve(ggml_backend_dev_count());
|
||||
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
|
||||
devs.push_back(ggml_backend_dev_get(i));
|
||||
}
|
||||
GGML_ASSERT(devs.size() >= 2);
|
||||
GGML_ASSERT(ggml_backend_dev_buffer_type(devs.back()) == ggml_backend_cpu_buffer_type());
|
||||
gpus.push_back(ggml_backend_meta_device(devs.data(), devs.size() - 1, llama_meta_device_get_tensor_split, nullptr));
|
||||
} else {
|
||||
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
|
||||
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
|
||||
switch (ggml_backend_dev_type(dev)) {
|
||||
case GGML_BACKEND_DEVICE_TYPE_CPU:
|
||||
case GGML_BACKEND_DEVICE_TYPE_ACCEL:
|
||||
// skip CPU backends since they are handled separately
|
||||
break;
|
||||
|
||||
case GGML_BACKEND_DEVICE_TYPE_GPU: {
|
||||
ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
|
||||
if (ggml_backend_reg_name(reg) == std::string("RPC")) {
|
||||
rpc_servers.push_back(dev);
|
||||
} else {
|
||||
// check if there is already a GPU with the same device id
|
||||
ggml_backend_dev_props props;
|
||||
ggml_backend_dev_get_props(dev, &props);
|
||||
auto it = std::find_if(gpus.begin(), gpus.end(), [&props](ggml_backend_dev_t d) {
|
||||
ggml_backend_dev_props d_props;
|
||||
ggml_backend_dev_get_props(d, &d_props);
|
||||
if (props.device_id && d_props.device_id) {
|
||||
return strcmp(props.device_id, d_props.device_id) == 0;
|
||||
}
|
||||
return false;
|
||||
});
|
||||
|
||||
if (it != gpus.end()) {
|
||||
LLAMA_LOG_INFO("%s: skipping device %s (%s) with id %s - already using device %s (%s) with the same id\n",
|
||||
__func__,
|
||||
ggml_backend_dev_name(dev), ggml_backend_dev_description(dev),
|
||||
props.device_id ? props.device_id : "unknown id",
|
||||
ggml_backend_dev_name(*it), ggml_backend_dev_description(*it));
|
||||
case GGML_BACKEND_DEVICE_TYPE_GPU: {
|
||||
ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
|
||||
if (ggml_backend_reg_name(reg) == std::string("RPC")) {
|
||||
rpc_servers.push_back(dev);
|
||||
} else {
|
||||
gpus.push_back(dev);
|
||||
}
|
||||
}
|
||||
break;
|
||||
}
|
||||
// check if there is already a GPU with the same device id
|
||||
ggml_backend_dev_props props;
|
||||
ggml_backend_dev_get_props(dev, &props);
|
||||
auto it = std::find_if(gpus.begin(), gpus.end(), [&props](ggml_backend_dev_t d) {
|
||||
ggml_backend_dev_props d_props;
|
||||
ggml_backend_dev_get_props(d, &d_props);
|
||||
if (props.device_id && d_props.device_id) {
|
||||
return strcmp(props.device_id, d_props.device_id) == 0;
|
||||
}
|
||||
return false;
|
||||
});
|
||||
|
||||
case GGML_BACKEND_DEVICE_TYPE_IGPU:
|
||||
igpus.push_back(dev);
|
||||
break;
|
||||
if (it != gpus.end()) {
|
||||
LLAMA_LOG_INFO("%s: skipping device %s (%s) with id %s - already using device %s (%s) with the same id\n",
|
||||
__func__,
|
||||
ggml_backend_dev_name(dev), ggml_backend_dev_description(dev),
|
||||
props.device_id ? props.device_id : "unknown id",
|
||||
ggml_backend_dev_name(*it), ggml_backend_dev_description(*it));
|
||||
} else {
|
||||
gpus.push_back(dev);
|
||||
}
|
||||
}
|
||||
break;
|
||||
}
|
||||
|
||||
case GGML_BACKEND_DEVICE_TYPE_IGPU:
|
||||
igpus.push_back(dev);
|
||||
break;
|
||||
case GGML_BACKEND_DEVICE_TYPE_META:
|
||||
GGML_ABORT("fatal error");
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -265,9 +265,15 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
|
||||
cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
|
||||
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp);
|
||||
ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
|
||||
1, chunk_size, n_chunks, g_diff_exp->ne[3]);
|
||||
|
||||
ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
|
||||
cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
|
||||
|
||||
ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
|
||||
cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
|
||||
|
||||
|
||||
// state to be updated per chunk
|
||||
ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
|
||||
@@ -322,9 +328,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
|
||||
: ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
|
||||
|
||||
// kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
|
||||
ggml_tensor * k_gdiff = ggml_cont(ctx0, get_slice_2d(ctx0, key_gdiff, chunk));
|
||||
ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
|
||||
//ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
|
||||
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, k_gdiff)));
|
||||
ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
|
||||
|
||||
// last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
|
||||
ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
|
||||
|
||||
@@ -169,20 +169,22 @@ static void init_tensor_kq_mask(ggml_tensor * tensor, float min = -1.0f, float m
|
||||
const int blck0 = 128;
|
||||
const int blck1 = 64;
|
||||
|
||||
// number of INF blocks
|
||||
const int n_inf_blocks = 0.1*(ne0*ne1*ne2*ne3)/(blck0*blck1);
|
||||
// number of INF/zero blocks
|
||||
const int n_inf_zero_blocks = 0.2*(ne0*ne1*ne2*ne3)/(blck0*blck1);
|
||||
|
||||
for (int b = 0; b < n_inf_blocks; b++) {
|
||||
for (int b = 0; b < n_inf_zero_blocks; b++) {
|
||||
const int p3 = (rd() % ne3);
|
||||
const int p2 = (rd() % ne2);
|
||||
const int p1 = (rd() % ne1);
|
||||
const int p0 = (rd() % ne0);
|
||||
|
||||
bool inf = rd() & 1;
|
||||
|
||||
for (int i1 = 0; i1 < blck1 && p1 + i1 < ne1; i1++) {
|
||||
const int idx = p3*ne2*ne1*ne0 + p2*ne1*ne0 + (p1 + i1)*ne0 + p0;
|
||||
|
||||
for (int i0 = 0; i0 < blck0 && p0 + i0 < ne0; i0++) {
|
||||
data_f32[idx + i0] = -INFINITY;
|
||||
data_f32[idx + i0] = inf ? -INFINITY : 0.0f;
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -8032,6 +8034,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
|
||||
for (int mode : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX, GGML_ROPE_TYPE_MROPE, GGML_ROPE_TYPE_IMROPE, GGML_ROPE_TYPE_VISION}) {
|
||||
for (bool ff : {false, true}) {
|
||||
test_cases.emplace_back(new test_rope(type, {128, 32, 2, 1}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 0, true, true));
|
||||
test_cases.emplace_back(new test_rope(type, {128, 32, 2, 1}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 1, true, true));
|
||||
test_cases.emplace_back(new test_rope(type, {128, 32, 2, 3}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 1, true, true));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -674,15 +674,12 @@ int main(int argc, char ** argv) {
|
||||
}
|
||||
}
|
||||
|
||||
for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
|
||||
int n_eval = (int) embd.size() - i;
|
||||
if (n_eval > params.n_batch) {
|
||||
n_eval = params.n_batch;
|
||||
}
|
||||
|
||||
if (!embd.empty()) {
|
||||
int n_eval = (int) embd.size();
|
||||
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
|
||||
|
||||
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
|
||||
GGML_ASSERT(n_eval <= params.n_batch);
|
||||
if (llama_decode(ctx, llama_batch_get_one(embd.data(), n_eval))) {
|
||||
LOG_ERR("%s : failed to eval\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
@@ -743,7 +740,7 @@ int main(int argc, char ** argv) {
|
||||
common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
|
||||
|
||||
++n_consumed;
|
||||
if ((int) embd.size() >= params.n_batch) {
|
||||
if ((int) embd.size() == params.n_batch) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -259,6 +259,8 @@ static const char * split_mode_str(llama_split_mode mode) {
|
||||
return "layer";
|
||||
case LLAMA_SPLIT_MODE_ROW:
|
||||
return "row";
|
||||
case LLAMA_SPLIT_MODE_TENSOR:
|
||||
return "tensor";
|
||||
default:
|
||||
GGML_ABORT("invalid split mode");
|
||||
}
|
||||
@@ -440,7 +442,7 @@ static void print_usage(int /* argc */, char ** argv) {
|
||||
join(cmd_params_defaults.n_gpu_layers, ",").c_str());
|
||||
printf(" -ncmoe, --n-cpu-moe <n> (default: %s)\n",
|
||||
join(cmd_params_defaults.n_cpu_moe, ",").c_str());
|
||||
printf(" -sm, --split-mode <none|layer|row> (default: %s)\n",
|
||||
printf(" -sm, --split-mode <none|layer|row|tensor> (default: %s)\n",
|
||||
join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
|
||||
printf(" -mg, --main-gpu <i> (default: %s)\n",
|
||||
join(cmd_params_defaults.main_gpu, ",").c_str());
|
||||
@@ -723,6 +725,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
|
||||
mode = LLAMA_SPLIT_MODE_LAYER;
|
||||
} else if (m == "row") {
|
||||
mode = LLAMA_SPLIT_MODE_ROW;
|
||||
} else if (m == "tensor") {
|
||||
mode = LLAMA_SPLIT_MODE_TENSOR;
|
||||
} else {
|
||||
invalid_param = true;
|
||||
break;
|
||||
@@ -1685,7 +1689,7 @@ struct markdown_printer : public printer {
|
||||
return 6;
|
||||
}
|
||||
if (field == "split_mode") {
|
||||
return 5;
|
||||
return 6;
|
||||
}
|
||||
if (field == "flash_attn") {
|
||||
return 2;
|
||||
|
||||
Vendored
+25
-3
@@ -3,9 +3,14 @@ license_add_file("cpp-httplib" "LICENSE")
|
||||
|
||||
find_package(Threads REQUIRED)
|
||||
|
||||
llama_add_compile_flags()
|
||||
|
||||
add_library(${TARGET} STATIC httplib.cpp httplib.h)
|
||||
if (NOT MSVC)
|
||||
# disable warnings in 3rd party code
|
||||
|
||||
# disable warnings in 3rd party code
|
||||
if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
|
||||
target_compile_options(${TARGET} PRIVATE /w)
|
||||
else()
|
||||
target_compile_options(${TARGET} PRIVATE -w)
|
||||
endif()
|
||||
|
||||
@@ -34,7 +39,7 @@ if (LLAMA_BUILD_BORINGSSL)
|
||||
set(FIPS OFF CACHE BOOL "Enable FIPS (BoringSSL)")
|
||||
|
||||
set(BORINGSSL_GIT "https://boringssl.googlesource.com/boringssl" CACHE STRING "BoringSSL git repository")
|
||||
set(BORINGSSL_VERSION "0.20251002.0" CACHE STRING "BoringSSL version")
|
||||
set(BORINGSSL_VERSION "0.20260204.0" CACHE STRING "BoringSSL version")
|
||||
|
||||
message(STATUS "Fetching BoringSSL version ${BORINGSSL_VERSION}")
|
||||
|
||||
@@ -146,6 +151,23 @@ elseif (LLAMA_OPENSSL)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
# disable warnings in 3rd party code
|
||||
if(LLAMA_BUILD_BORINGSSL OR LLAMA_BUILD_LIBRESSL)
|
||||
if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
|
||||
target_compile_options(ssl PRIVATE /w)
|
||||
target_compile_options(crypto PRIVATE /w)
|
||||
if(LLAMA_BUILD_BORINGSSL)
|
||||
target_compile_options(fipsmodule PRIVATE /w)
|
||||
endif()
|
||||
else()
|
||||
target_compile_options(ssl PRIVATE -w)
|
||||
target_compile_options(crypto PRIVATE -w)
|
||||
if(LLAMA_BUILD_BORINGSSL)
|
||||
target_compile_options(fipsmodule PRIVATE -w)
|
||||
endif()
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if (CPPHTTPLIB_OPENSSL_SUPPORT)
|
||||
target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_OPENSSL_SUPPORT) # used in server.cpp
|
||||
if (APPLE AND CMAKE_SYSTEM_NAME STREQUAL "Darwin")
|
||||
|
||||
Vendored
+190
-48
@@ -117,6 +117,8 @@ time_t parse_http_date(const std::string &date_str) {
|
||||
|
||||
#ifdef _WIN32
|
||||
return _mkgmtime(&tm_buf);
|
||||
#elif defined _AIX
|
||||
return mktime(&tm_buf);
|
||||
#else
|
||||
return timegm(&tm_buf);
|
||||
#endif
|
||||
@@ -1376,7 +1378,7 @@ int getaddrinfo_with_timeout(const char *node, const char *service,
|
||||
|
||||
// Allocate on the heap, so the resolver thread can keep using the data.
|
||||
auto state = std::make_shared<GetAddrInfoState>();
|
||||
state->node = node;
|
||||
if (node) { state->node = node; }
|
||||
state->service = service;
|
||||
state->hints = *hints;
|
||||
|
||||
@@ -2896,10 +2898,20 @@ bool parse_range_header(const std::string &s, Ranges &ranges) try {
|
||||
return;
|
||||
}
|
||||
|
||||
const auto first =
|
||||
static_cast<ssize_t>(lhs.empty() ? -1 : std::stoll(lhs));
|
||||
const auto last =
|
||||
static_cast<ssize_t>(rhs.empty() ? -1 : std::stoll(rhs));
|
||||
ssize_t first = -1;
|
||||
if (!lhs.empty()) {
|
||||
ssize_t v;
|
||||
auto res = detail::from_chars(lhs.data(), lhs.data() + lhs.size(), v);
|
||||
if (res.ec == std::errc{}) { first = v; }
|
||||
}
|
||||
|
||||
ssize_t last = -1;
|
||||
if (!rhs.empty()) {
|
||||
ssize_t v;
|
||||
auto res = detail::from_chars(rhs.data(), rhs.data() + rhs.size(), v);
|
||||
if (res.ec == std::errc{}) { last = v; }
|
||||
}
|
||||
|
||||
if ((first == -1 && last == -1) ||
|
||||
(first != -1 && last != -1 && first > last)) {
|
||||
all_valid_ranges = false;
|
||||
@@ -2974,25 +2986,17 @@ bool parse_accept_header(const std::string &s,
|
||||
return;
|
||||
}
|
||||
|
||||
#ifdef CPPHTTPLIB_NO_EXCEPTIONS
|
||||
{
|
||||
std::istringstream iss(quality_str);
|
||||
iss >> accept_entry.quality;
|
||||
|
||||
// Check if conversion was successful and entire string was consumed
|
||||
if (iss.fail() || !iss.eof()) {
|
||||
double v = 0.0;
|
||||
auto res = detail::from_chars(
|
||||
quality_str.data(), quality_str.data() + quality_str.size(), v);
|
||||
if (res.ec == std::errc{}) {
|
||||
accept_entry.quality = v;
|
||||
} else {
|
||||
has_invalid_entry = true;
|
||||
return;
|
||||
}
|
||||
}
|
||||
#else
|
||||
try {
|
||||
accept_entry.quality = std::stod(quality_str);
|
||||
} catch (...) {
|
||||
has_invalid_entry = true;
|
||||
return;
|
||||
}
|
||||
#endif
|
||||
// Check if quality is in valid range [0.0, 1.0]
|
||||
if (accept_entry.quality < 0.0 || accept_entry.quality > 1.0) {
|
||||
has_invalid_entry = true;
|
||||
@@ -5570,13 +5574,26 @@ bool Server::read_content(Stream &strm, Request &req, Response &res) {
|
||||
strm, req, res,
|
||||
// Regular
|
||||
[&](const char *buf, size_t n) {
|
||||
// Prevent arithmetic overflow when checking sizes.
|
||||
// Avoid computing (req.body.size() + n) directly because
|
||||
// adding two unsigned `size_t` values can wrap around and
|
||||
// produce a small result instead of indicating overflow.
|
||||
// Instead, check using subtraction: ensure `n` does not
|
||||
// exceed the remaining capacity `max_size() - size()`.
|
||||
if (req.body.size() >= req.body.max_size() ||
|
||||
n > req.body.max_size() - req.body.size()) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// Limit decompressed body size to payload_max_length_ to protect
|
||||
// against "zip bomb" attacks where a small compressed payload
|
||||
// decompresses to a massive size.
|
||||
if (req.body.size() + n > payload_max_length_ ||
|
||||
req.body.size() + n > req.body.max_size()) {
|
||||
if (payload_max_length_ > 0 &&
|
||||
(req.body.size() >= payload_max_length_ ||
|
||||
n > payload_max_length_ - req.body.size())) {
|
||||
return false;
|
||||
}
|
||||
|
||||
req.body.append(buf, n);
|
||||
return true;
|
||||
},
|
||||
@@ -5666,22 +5683,29 @@ bool Server::read_content_core(
|
||||
// oversized request and fail early (causing connection close). For SSL
|
||||
// builds we cannot reliably peek the decrypted application bytes, so keep
|
||||
// the original behaviour.
|
||||
#if !defined(CPPHTTPLIB_OPENSSL_SUPPORT) && !defined(_WIN32)
|
||||
#if !defined(CPPHTTPLIB_OPENSSL_SUPPORT)
|
||||
if (!req.has_header("Content-Length") &&
|
||||
!detail::is_chunked_transfer_encoding(req.headers)) {
|
||||
socket_t s = strm.socket();
|
||||
if (s != INVALID_SOCKET) {
|
||||
// Peek up to payload_max_length_ + 1 bytes. If more than
|
||||
// payload_max_length_ bytes are pending, reject the request.
|
||||
size_t to_peek =
|
||||
(payload_max_length_ > 0)
|
||||
? (std::min)(payload_max_length_ + 1, static_cast<size_t>(4096))
|
||||
: 1;
|
||||
std::vector<char> peekbuf(to_peek);
|
||||
ssize_t n = ::recv(s, peekbuf.data(), to_peek, MSG_PEEK);
|
||||
if (n > 0 && static_cast<size_t>(n) > payload_max_length_) {
|
||||
// Indicate failure so connection will be closed.
|
||||
return false;
|
||||
// Only peek if payload_max_length is set to a finite value
|
||||
if (payload_max_length_ > 0 &&
|
||||
payload_max_length_ < (std::numeric_limits<size_t>::max)()) {
|
||||
socket_t s = strm.socket();
|
||||
if (s != INVALID_SOCKET) {
|
||||
// Peek to check if there is any pending data
|
||||
char peekbuf[1];
|
||||
ssize_t n = ::recv(s, peekbuf, 1, MSG_PEEK);
|
||||
if (n > 0) {
|
||||
// There is data, so read it with payload limit enforcement
|
||||
auto result = detail::read_content_without_length(
|
||||
strm, payload_max_length_, out);
|
||||
if (result == detail::ReadContentResult::PayloadTooLarge) {
|
||||
res.status = StatusCode::PayloadTooLarge_413;
|
||||
return false;
|
||||
} else if (result != detail::ReadContentResult::Success) {
|
||||
return false;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
}
|
||||
}
|
||||
return true;
|
||||
@@ -6656,7 +6680,8 @@ void ClientImpl::close_socket(Socket &socket) {
|
||||
}
|
||||
|
||||
bool ClientImpl::read_response_line(Stream &strm, const Request &req,
|
||||
Response &res) const {
|
||||
Response &res,
|
||||
bool skip_100_continue) const {
|
||||
std::array<char, 2048> buf{};
|
||||
|
||||
detail::stream_line_reader line_reader(strm, buf.data(), buf.size());
|
||||
@@ -6677,8 +6702,8 @@ bool ClientImpl::read_response_line(Stream &strm, const Request &req,
|
||||
res.status = std::stoi(std::string(m[2]));
|
||||
res.reason = std::string(m[3]);
|
||||
|
||||
// Ignore '100 Continue'
|
||||
while (res.status == StatusCode::Continue_100) {
|
||||
// Ignore '100 Continue' (only when not using Expect: 100-continue explicitly)
|
||||
while (skip_100_continue && res.status == StatusCode::Continue_100) {
|
||||
if (!line_reader.getline()) { return false; } // CRLF
|
||||
if (!line_reader.getline()) { return false; } // next response line
|
||||
|
||||
@@ -7463,7 +7488,8 @@ bool ClientImpl::write_content_with_provider(Stream &strm,
|
||||
}
|
||||
|
||||
bool ClientImpl::write_request(Stream &strm, Request &req,
|
||||
bool close_connection, Error &error) {
|
||||
bool close_connection, Error &error,
|
||||
bool skip_body) {
|
||||
// Prepare additional headers
|
||||
if (close_connection) {
|
||||
if (!req.has_header("Connection")) {
|
||||
@@ -7582,7 +7608,59 @@ bool ClientImpl::write_request(Stream &strm, Request &req,
|
||||
}
|
||||
}
|
||||
|
||||
// After sending request line and headers, wait briefly for an early server
|
||||
// response (e.g. 4xx) and avoid sending a potentially large request body
|
||||
// unnecessarily. This workaround is only enabled on Windows because Unix
|
||||
// platforms surface write errors (EPIPE) earlier; on Windows kernel send
|
||||
// buffering can accept large writes even when the peer already responded.
|
||||
// Check the stream first (which covers SSL via `is_readable()`), then
|
||||
// fall back to select on the socket. Only perform the wait for very large
|
||||
// request bodies to avoid interfering with normal small requests and
|
||||
// reduce side-effects. Poll briefly (up to 50ms as default) for an early
|
||||
// response. Skip this check when using Expect: 100-continue, as the protocol
|
||||
// handles early responses properly.
|
||||
#if defined(_WIN32)
|
||||
if (!skip_body &&
|
||||
req.body.size() > CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_THRESHOLD &&
|
||||
req.path.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) {
|
||||
auto start = std::chrono::high_resolution_clock::now();
|
||||
|
||||
for (;;) {
|
||||
// Prefer socket-level readiness to avoid SSL_pending() false-positives
|
||||
// from SSL internals. If the underlying socket is readable, assume an
|
||||
// early response may be present.
|
||||
auto sock = strm.socket();
|
||||
if (sock != INVALID_SOCKET && detail::select_read(sock, 0, 0) > 0) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// Fallback to stream-level check for non-socket streams or when the
|
||||
// socket isn't reporting readable. Avoid using `is_readable()` for
|
||||
// SSL, since `SSL_pending()` may report buffered records that do not
|
||||
// indicate a complete application-level response yet.
|
||||
if (!is_ssl() && strm.is_readable()) { return false; }
|
||||
|
||||
auto now = std::chrono::high_resolution_clock::now();
|
||||
auto elapsed =
|
||||
std::chrono::duration_cast<std::chrono::milliseconds>(now - start)
|
||||
.count();
|
||||
if (elapsed >= CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_TIMEOUT_MSECOND) {
|
||||
break;
|
||||
}
|
||||
|
||||
std::this_thread::sleep_for(std::chrono::milliseconds(1));
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
// Body
|
||||
if (skip_body) { return true; }
|
||||
|
||||
return write_request_body(strm, req, error);
|
||||
}
|
||||
|
||||
bool ClientImpl::write_request_body(Stream &strm, Request &req,
|
||||
Error &error) {
|
||||
if (req.body.empty()) {
|
||||
return write_content_with_provider(strm, req, error);
|
||||
}
|
||||
@@ -7758,8 +7836,20 @@ void ClientImpl::output_error_log(const Error &err,
|
||||
bool ClientImpl::process_request(Stream &strm, Request &req,
|
||||
Response &res, bool close_connection,
|
||||
Error &error) {
|
||||
// Send request
|
||||
if (!write_request(strm, req, close_connection, error)) { return false; }
|
||||
// Auto-add Expect: 100-continue for large bodies
|
||||
if (CPPHTTPLIB_EXPECT_100_THRESHOLD > 0 && !req.has_header("Expect")) {
|
||||
auto body_size = req.body.empty() ? req.content_length_ : req.body.size();
|
||||
if (body_size >= CPPHTTPLIB_EXPECT_100_THRESHOLD) {
|
||||
req.set_header("Expect", "100-continue");
|
||||
}
|
||||
}
|
||||
|
||||
// Check for Expect: 100-continue
|
||||
auto expect_100_continue = req.get_header_value("Expect") == "100-continue";
|
||||
|
||||
// Send request (skip body if using Expect: 100-continue)
|
||||
auto write_request_success =
|
||||
write_request(strm, req, close_connection, error, expect_100_continue);
|
||||
|
||||
#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
|
||||
if (is_ssl()) {
|
||||
@@ -7774,14 +7864,48 @@ bool ClientImpl::process_request(Stream &strm, Request &req,
|
||||
}
|
||||
#endif
|
||||
|
||||
// Handle Expect: 100-continue with timeout
|
||||
if (expect_100_continue && CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND > 0) {
|
||||
time_t sec = CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND / 1000;
|
||||
time_t usec = (CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND % 1000) * 1000;
|
||||
auto ret = detail::select_read(strm.socket(), sec, usec);
|
||||
if (ret <= 0) {
|
||||
// Timeout or error: send body anyway (server didn't respond in time)
|
||||
if (!write_request_body(strm, req, error)) { return false; }
|
||||
expect_100_continue = false; // Switch to normal response handling
|
||||
}
|
||||
}
|
||||
|
||||
// Receive response and headers
|
||||
if (!read_response_line(strm, req, res) ||
|
||||
// When using Expect: 100-continue, don't auto-skip `100 Continue` response
|
||||
if (!read_response_line(strm, req, res, !expect_100_continue) ||
|
||||
!detail::read_headers(strm, res.headers)) {
|
||||
error = Error::Read;
|
||||
if (write_request_success) { error = Error::Read; }
|
||||
output_error_log(error, &req);
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!write_request_success) { return false; }
|
||||
|
||||
// Handle Expect: 100-continue response
|
||||
if (expect_100_continue) {
|
||||
if (res.status == StatusCode::Continue_100) {
|
||||
// Server accepted, send the body
|
||||
if (!write_request_body(strm, req, error)) { return false; }
|
||||
|
||||
// Read the actual response
|
||||
res.headers.clear();
|
||||
res.body.clear();
|
||||
if (!read_response_line(strm, req, res) ||
|
||||
!detail::read_headers(strm, res.headers)) {
|
||||
error = Error::Read;
|
||||
output_error_log(error, &req);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
// If not 100 Continue, server returned an error; proceed with that response
|
||||
}
|
||||
|
||||
// Body
|
||||
if ((res.status != StatusCode::NoContent_204) && req.method != "HEAD" &&
|
||||
req.method != "CONNECT") {
|
||||
@@ -9543,7 +9667,7 @@ bool SSLClient::load_certs() {
|
||||
last_openssl_error_ = ERR_get_error();
|
||||
ret = false;
|
||||
}
|
||||
} else {
|
||||
} else if (!ca_cert_store_) {
|
||||
auto loaded = false;
|
||||
#ifdef _WIN32
|
||||
loaded =
|
||||
@@ -9790,7 +9914,11 @@ bool SSLClient::verify_host_with_common_name(X509 *server_cert) const {
|
||||
|
||||
bool SSLClient::check_host_name(const char *pattern,
|
||||
size_t pattern_len) const {
|
||||
if (host_.size() == pattern_len && host_ == pattern) { return true; }
|
||||
// Exact match (case-insensitive)
|
||||
if (host_.size() == pattern_len &&
|
||||
detail::case_ignore::equal(host_, std::string(pattern, pattern_len))) {
|
||||
return true;
|
||||
}
|
||||
|
||||
// Wildcard match
|
||||
// https://bugs.launchpad.net/ubuntu/+source/firefox-3.0/+bug/376484
|
||||
@@ -9805,9 +9933,23 @@ bool SSLClient::check_host_name(const char *pattern,
|
||||
auto itr = pattern_components.begin();
|
||||
for (const auto &h : host_components_) {
|
||||
auto &p = *itr;
|
||||
if (p != h && p != "*") {
|
||||
auto partial_match = (p.size() > 0 && p[p.size() - 1] == '*' &&
|
||||
!p.compare(0, p.size() - 1, h));
|
||||
if (!httplib::detail::case_ignore::equal(p, h) && p != "*") {
|
||||
bool partial_match = false;
|
||||
if (!p.empty() && p[p.size() - 1] == '*') {
|
||||
const auto prefix_length = p.size() - 1;
|
||||
if (prefix_length == 0) {
|
||||
partial_match = true;
|
||||
} else if (h.size() >= prefix_length) {
|
||||
partial_match =
|
||||
std::equal(p.begin(),
|
||||
p.begin() + static_cast<std::string::difference_type>(
|
||||
prefix_length),
|
||||
h.begin(), [](const char ca, const char cb) {
|
||||
return httplib::detail::case_ignore::to_lower(ca) ==
|
||||
httplib::detail::case_ignore::to_lower(cb);
|
||||
});
|
||||
}
|
||||
}
|
||||
if (!partial_match) { return false; }
|
||||
}
|
||||
++itr;
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user