graph : add comments

ggml-ci
graph : move non-context related logic to llm_build_context
2026-07-16 01:15:58 +02:00 · 2025-02-28 21:13:08 +02:00 · 2025-02-28 20:36:25 +02:00 · 2025-02-28 18:01:25 +02:00 · 2025-02-28 16:30:41 +02:00 · 2025-02-28 10:51:17 +02:00
139 changed files with 20837 additions and 13090 deletions
@@ -173,7 +173,15 @@ jobs:
          name: llama-bin-macos-x64.zip

  ubuntu-cpu-cmake:
-    runs-on: ubuntu-22.04
+    strategy:
+      matrix:
+        include:
+          - build: 'x64'
+            os: ubuntu-22.04
+          - build: 'arm64'
+            os: ubuntu-22.04-arm
+
+    runs-on: ${{ matrix.os }}

    steps:
      - name: Clone
@@ -239,14 +247,14 @@ jobs:
        run: |
          cp LICENSE ./build/bin/
          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/*
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/*

      - name: Upload artifacts
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
        uses: actions/upload-artifact@v4
        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
-          name: llama-bin-ubuntu-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip
+          name: llama-bin-ubuntu-${{ matrix.build }}.zip

  ubuntu-latest-cmake-sanitizer:
    runs-on: ubuntu-latest
@@ -680,6 +680,10 @@ ifdef GGML_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN)
 endif # GGML_CUDA_CCBIN

+ifdef GGML_CUDA_NO_FA
+	MK_NVCCFLAGS += -DGGML_CUDA_NO_FA
+endif # GGML_CUDA_NO_FA
+
 ifdef GGML_CUDA_FA_ALL_QUANTS
 	MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
 endif # GGML_CUDA_FA_ALL_QUANTS
@@ -800,6 +804,10 @@ ifdef GGML_CUDA_NO_PEER_COPY
 	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
 endif # GGML_CUDA_NO_PEER_COPY

+ifdef GGML_CUDA_NO_FA
+	HIPFLAGS += -DGGML_CUDA_NO_FA
+endif # GGML_CUDA_NO_FA
+
 	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
 	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
 	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
@@ -876,6 +884,10 @@ ifdef GGML_CUDA_NO_PEER_COPY
 	MUSAFLAGS += -DGGML_CUDA_NO_PEER_COPY
 endif # GGML_CUDA_NO_PEER_COPY

+ifdef GGML_CUDA_NO_FA
+	MUSAFLAGS += -DGGML_CUDA_NO_FA
+endif # GGML_CUDA_NO_FA
+
 ifdef GGML_CUDA_FA_ALL_QUANTS
 	MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
 endif # GGML_CUDA_FA_ALL_QUANTS
@@ -219,7 +219,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
 - [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
 - [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale
-
+- [llmaz](https://github.com/InftyAI/llmaz) - ☸️ Easy, advanced inference platform for large language models on Kubernetes.
 </details>

 <details>
@@ -951,8 +951,8 @@ struct common_init_result common_init_from_params(common_params & params) {
        return iparams;
    }

-    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
        params.ctx_shift = false;
    }

@@ -1056,7 +1056,7 @@ struct common_init_result common_init_from_params(common_params & params) {
        if (llama_model_has_decoder(model)) {
            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
        }
-        llama_kv_cache_clear(lctx);
+        llama_kv_self_clear(lctx);
        llama_synchronize(lctx);
        llama_perf_context_reset(lctx);
    }
@@ -172,7 +172,7 @@ llama_tokens common_speculative_gen_draft(
    result.reserve(params.n_draft);

    if (reuse_n == 0) {
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        prompt.clear();
    } else {
@@ -191,14 +191,14 @@ llama_tokens common_speculative_gen_draft(
        }

        if (reuse_i > 0) {
-            llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
+            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);

            prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
        }

        if (reuse_n < (int) prompt.size()) {
-            llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1);
+            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);

            prompt.erase(prompt.begin() + reuse_n, prompt.end());
        }
@@ -42,6 +42,16 @@ The following release is verified with good quality:

 ## News

+- 2025.2
+  - Optimize MUL_MAT Q4_0 on Intel GPU for all dGPUs and built-in GPUs since MTL. Increase the performance of LLM (llama-2-7b.Q4_0.gguf) 21%-87% on Intel GPUs (MTL, ARL-H, Arc, Flex, PVC).
+    |GPU|Base tokens/s|Increased tokens/s|Percent|
+    |-|-|-|-|
+    |PVC 1550|39|73|+87%|
+    |Flex 170|39|50|+28%|
+    |Arc770|42|55|+30%|
+    |MTL|13|16|+23%|
+    |ARL-H|14|17|+21%|
+
 - 2024.11
  - Use syclcompat to improve the performance on some platforms. This requires to use oneAPI 2025.0 or newer.

@@ -97,8 +107,8 @@ SYCL backend supports Intel GPU Family:
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
-| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake    |
+| Intel iGPU                    | Support | iGPU in 13700k,iGPU in 13400, i5-1250P, i7-1260P, i7-1165G7 |

 *Notes:*

@@ -660,8 +670,10 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
+| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |

+
 ## Known Issues

 - `Split-mode:[row]` is not supported.
@@ -0,0 +1,390 @@
+# Function Calling
+
+[chat.h](../common/chat.h) (https://github.com/ggml-org/llama.cpp/pull/9639) adds support for [OpenAI-style function calling](https://platform.openai.com/docs/guides/function-calling) and is used in:
+- `llama-server` when started w/ `--jinja` flag
+- `llama-cli` (WIP: https://github.com/ggml-org/llama.cpp/pull/11556)
+
+## Universal support w/ Native & Generic handlers
+
+Function calling is supported for all models (see https://github.com/ggml-org/llama.cpp/pull/9639):
+
+- Native tool call formats supported:
+  - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2
+  - Functionary v3.1 / v3.2
+  - Hermes 2/3, Qwen 2.5
+  - Qwen 2.5 Coder (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
+  - Mistral Nemo
+  - Firefunction v2
+  - Command R7B
+  - DeepSeek R1 (WIP / seems reluctant to call any tools?)
+
+- Generic tool call is supported when the template isn't recognized by native format handlers (you'll see `Chat format: Generic` in the logs).
+  - Use `--chat-template-file` to override the template when appropriate (see examples below)
+  - Generic support may consume more tokens and be less efficient than a model's native format.
+
+<details>
+<summary>Show some common templates and which format handler they use</summary>
+
+| Template | Format |
+|----------|--------|
+| Almawave-Velvet-14B.jinja | Hermes 2 Pro |
+| AtlaAI-Selene-1-Mini-Llama-3.1-8B.jinja | Llama 3.x |
+| CohereForAI-aya-expanse-8b.jinja | Generic |
+| CohereForAI-c4ai-command-r-plus-default.jinja | Generic |
+| CohereForAI-c4ai-command-r-plus-rag.jinja | Generic |
+| CohereForAI-c4ai-command-r-plus-tool_use.jinja | Generic |
+| CohereForAI-c4ai-command-r7b-12-2024-default.jinja | Command R7B (extract reasoning) |
+| CohereForAI-c4ai-command-r7b-12-2024-rag.jinja | Command R7B (extract reasoning) |
+| CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja | Command R7B (extract reasoning) |
+| CohereForAI-c4ai-command-r7b-12-2024.jinja | Generic |
+| DavieLion-Llama-3.2-1B-SPIN-iter3.jinja | Generic |
+| Delta-Vector-Rei-12B.jinja | Mistral Nemo |
+| EpistemeAI-Mistral-Nemo-Instruct-12B-Philosophy-Math.jinja | Mistral Nemo |
+| FlofloB-83k_continued_pretraining_Qwen2.5-0.5B-Instruct_Unsloth_merged_16bit.jinja | Hermes 2 Pro |
+| FlofloB-test_continued_pretraining_Phi-3-mini-4k-instruct_Unsloth_merged_16bit.jinja | Generic |
+| HelpingAI-HAI-SER.jinja | Generic |
+| HuggingFaceTB-SmolLM2-1.7B-Instruct.jinja | Generic |
+| HuggingFaceTB-SmolLM2-135M-Instruct.jinja | Generic |
+| HuggingFaceTB-SmolLM2-360M-Instruct.jinja | Generic |
+| INSAIT-Institute-BgGPT-Gemma-2-27B-IT-v1.0.jinja | Generic |
+| Ihor-Text2Graph-R1-Qwen2.5-0.5b.jinja | Hermes 2 Pro |
+| Infinigence-Megrez-3B-Instruct.jinja | Generic |
+| Josephgflowers-TinyLlama_v1.1_math_code-world-test-1.jinja | Generic |
+| LGAI-EXAONE-EXAONE-3.5-2.4B-Instruct.jinja | Generic |
+| LGAI-EXAONE-EXAONE-3.5-7.8B-Instruct.jinja | Generic |
+| LatitudeGames-Wayfarer-12B.jinja | Generic |
+| Magpie-Align-Llama-3-8B-Magpie-Align-v0.1.jinja | Generic |
+| Magpie-Align-Llama-3.1-8B-Magpie-Align-v0.1.jinja | Generic |
+| MaziyarPanahi-calme-3.2-instruct-78b.jinja | Generic |
+| MiniMaxAI-MiniMax-Text-01.jinja | Generic |
+| MiniMaxAI-MiniMax-VL-01.jinja | Generic |
+| NaniDAO-deepseek-r1-qwen-2.5-32B-ablated.jinja | DeepSeek R1 (extract reasoning) |
+| NexaAIDev-Octopus-v2.jinja | Generic |
+| NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | Generic |
+| NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | Hermes 2 Pro |
+| NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | Generic |
+| NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | Hermes 2 Pro |
+| NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | Generic |
+| NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | Hermes 2 Pro |
+| NovaSky-AI-Sky-T1-32B-Flash.jinja | Hermes 2 Pro |
+| NovaSky-AI-Sky-T1-32B-Preview.jinja | Hermes 2 Pro |
+| OnlyCheeini-greesychat-turbo.jinja | Generic |
+| Orenguteng-Llama-3.1-8B-Lexi-Uncensored-V2.jinja | Llama 3.x |
+| OrionStarAI-Orion-14B-Chat.jinja | Generic |
+| PowerInfer-SmallThinker-3B-Preview.jinja | Generic |
+| PrimeIntellect-INTELLECT-1-Instruct.jinja | Generic |
+| Qwen-QVQ-72B-Preview.jinja | Generic |
+| Qwen-QwQ-32B-Preview.jinja | Hermes 2 Pro |
+| Qwen-Qwen1.5-7B-Chat.jinja | Generic |
+| Qwen-Qwen2-7B-Instruct.jinja | Generic |
+| Qwen-Qwen2-VL-72B-Instruct.jinja | Generic |
+| Qwen-Qwen2-VL-7B-Instruct.jinja | Generic |
+| Qwen-Qwen2.5-0.5B.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-1.5B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-14B-Instruct-1M.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-14B.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-32B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-32B.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-3B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-72B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-7B-Instruct-1M.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-7B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-7B.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-Coder-32B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-Coder-7B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-Math-1.5B.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-Math-7B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-VL-3B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-VL-72B-Instruct.jinja | Hermes 2 Pro |
+| Qwen-Qwen2.5-VL-7B-Instruct.jinja | Hermes 2 Pro |
+| RWKV-Red-Team-ARWKV-7B-Preview-0.1.jinja | Hermes 2 Pro |
+| SakanaAI-TinySwallow-1.5B-Instruct.jinja | Hermes 2 Pro |
+| SakanaAI-TinySwallow-1.5B.jinja | Hermes 2 Pro |
+| Sao10K-70B-L3.3-Cirrus-x1.jinja | Llama 3.x |
+| SentientAGI-Dobby-Mini-Leashed-Llama-3.1-8B.jinja | Llama 3.x |
+| SentientAGI-Dobby-Mini-Unhinged-Llama-3.1-8B.jinja | Llama 3.x |
+| Steelskull-L3.3-Damascus-R1.jinja | Llama 3.x |
+| Steelskull-L3.3-MS-Nevoria-70b.jinja | Llama 3.x |
+| Steelskull-L3.3-Nevoria-R1-70b.jinja | Llama 3.x |
+| THUDM-glm-4-9b-chat.jinja | Generic |
+| THUDM-glm-edge-1.5b-chat.jinja | Generic |
+| Tarek07-Progenitor-V1.1-LLaMa-70B.jinja | Llama 3.x |
+| TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | Generic |
+| TinyLlama-TinyLlama-1.1B-Chat-v1.0.jinja | Generic |
+| UCLA-AGI-Mistral7B-PairRM-SPPO-Iter3.jinja | Generic |
+| ValiantLabs-Llama3.1-8B-Enigma.jinja | Llama 3.x |
+| abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | Generic |
+| ai21labs-AI21-Jamba-1.5-Large.jinja | Generic |
+| allenai-Llama-3.1-Tulu-3-405B-SFT.jinja | Generic |
+| allenai-Llama-3.1-Tulu-3-405B.jinja | Generic |
+| allenai-Llama-3.1-Tulu-3-8B.jinja | Generic |
+| arcee-ai-Virtuoso-Lite.jinja | Hermes 2 Pro |
+| arcee-ai-Virtuoso-Medium-v2.jinja | Hermes 2 Pro |
+| arcee-ai-Virtuoso-Small-v2.jinja | Hermes 2 Pro |
+| avemio-GRAG-NEMO-12B-ORPO-HESSIAN-AI.jinja | Generic |
+| bespokelabs-Bespoke-Stratos-7B.jinja | Hermes 2 Pro |
+| bfuzzy1-acheron-m1a-llama.jinja | Generic |
+| bofenghuang-vigogne-2-70b-chat.jinja | Generic |
+| bytedance-research-UI-TARS-72B-DPO.jinja | Generic |
+| bytedance-research-UI-TARS-7B-DPO.jinja | Generic |
+| bytedance-research-UI-TARS-7B-SFT.jinja | Generic |
+| carsenk-phi3.5_mini_exp_825_uncensored.jinja | Generic |
+| cyberagent-DeepSeek-R1-Distill-Qwen-14B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
+| cyberagent-DeepSeek-R1-Distill-Qwen-32B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
+| databricks-dbrx-instruct.jinja | Generic |
+| deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | Generic |
+| deepseek-ai-DeepSeek-Coder-V2-Lite-Base.jinja | Generic |
+| deepseek-ai-DeepSeek-Coder-V2-Lite-Instruct.jinja | Generic |
+| deepseek-ai-DeepSeek-R1-Distill-Llama-70B.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-R1-Distill-Qwen-14B.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-R1-Zero.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-V2-Lite.jinja | Generic |
+| deepseek-ai-DeepSeek-V2.5.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-DeepSeek-V3.jinja | DeepSeek R1 (extract reasoning) |
+| deepseek-ai-deepseek-coder-33b-instruct.jinja | Generic |
+| deepseek-ai-deepseek-coder-6.7b-instruct.jinja | Generic |
+| deepseek-ai-deepseek-coder-7b-instruct-v1.5.jinja | Generic |
+| deepseek-ai-deepseek-llm-67b-chat.jinja | Generic |
+| deepseek-ai-deepseek-llm-7b-chat.jinja | Generic |
+| dicta-il-dictalm2.0-instruct.jinja | Generic |
+| ehristoforu-Falcon3-8B-Franken-Basestruct.jinja | Hermes 2 Pro |
+| fireworks-ai-llama-3-firefunction-v2.jinja | FireFunction v2 |
+| godlikehhd-alpaca_data_sampled_ifd_new_5200.jinja | Hermes 2 Pro |
+| godlikehhd-alpaca_data_score_max_0.7_2600.jinja | Hermes 2 Pro |
+| google-gemma-2-27b-it.jinja | Generic |
+| google-gemma-2-2b-it.jinja | Generic |
+| google-gemma-2-2b-jpn-it.jinja | Generic |
+| google-gemma-7b-it.jinja | Generic |
+| huihui-ai-DeepSeek-R1-Distill-Llama-70B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
+| huihui-ai-DeepSeek-R1-Distill-Llama-8B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
+| huihui-ai-DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
+| huihui-ai-DeepSeek-R1-Distill-Qwen-32B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
+| huihui-ai-DeepSeek-R1-Distill-Qwen-7B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
+| huihui-ai-Qwen2.5-14B-Instruct-1M-abliterated.jinja | Hermes 2 Pro |
+| ibm-granite-granite-3.1-8b-instruct.jinja | Generic |
+| indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | Generic |
+| inflatebot-MN-12B-Mag-Mell-R1.jinja | Generic |
+| jinaai-ReaderLM-v2.jinja | Generic |
+| kms7530-chemeng_qwen-math-7b_24_1_100_1_nonmath.jinja | Hermes 2 Pro |
+| knifeayumu-Cydonia-v1.3-Magnum-v4-22B.jinja | Mistral Nemo |
+| langgptai-qwen1.5-7b-chat-sa-v0.1.jinja | Generic |
+| lightblue-DeepSeek-R1-Distill-Qwen-7B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
+| mattshumer-Reflection-Llama-3.1-70B.jinja | Generic |
+| meetkai-functionary-medium-v3.1.jinja | Functionary v3.1 Llama 3.1 |
+| meetkai-functionary-medium-v3.2.jinja | Functionary v3.2 |
+| meta-llama-Llama-2-7b-chat-hf.jinja | Generic |
+| meta-llama-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
+| meta-llama-Llama-3.2-11B-Vision-Instruct.jinja | Llama 3.x |
+| meta-llama-Llama-3.2-1B-Instruct.jinja | Llama 3.x |
+| meta-llama-Llama-3.2-3B-Instruct.jinja | Llama 3.x |
+| meta-llama-Llama-3.3-70B-Instruct.jinja | Llama 3.x |
+| meta-llama-Meta-Llama-3-8B-Instruct.jinja | Generic |
+| meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
+| microsoft-Phi-3-medium-4k-instruct.jinja | Generic |
+| microsoft-Phi-3-mini-4k-instruct.jinja | Generic |
+| microsoft-Phi-3-small-8k-instruct.jinja | Generic |
+| microsoft-Phi-3.5-mini-instruct.jinja | Generic |
+| microsoft-Phi-3.5-vision-instruct.jinja | Generic |
+| microsoft-phi-4.jinja | Generic |
+| migtissera-Tess-3-Mistral-Nemo-12B.jinja | Generic |
+| ministral-Ministral-3b-instruct.jinja | Generic |
+| mistralai-Codestral-22B-v0.1.jinja | Generic |
+| mistralai-Mistral-7B-Instruct-v0.1.jinja | Generic |
+| mistralai-Mistral-7B-Instruct-v0.2.jinja | Generic |
+| mistralai-Mistral-7B-Instruct-v0.3.jinja | Mistral Nemo |
+| mistralai-Mistral-Large-Instruct-2407.jinja | Mistral Nemo |
+| mistralai-Mistral-Large-Instruct-2411.jinja | Generic |
+| mistralai-Mistral-Nemo-Instruct-2407.jinja | Mistral Nemo |
+| mistralai-Mistral-Small-24B-Instruct-2501.jinja | Generic |
+| mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | Generic |
+| mkurman-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
+| mlabonne-AlphaMonarch-7B.jinja | Generic |
+| mlx-community-Josiefied-Qwen2.5-0.5B-Instruct-abliterated-v1-float32.jinja | Hermes 2 Pro |
+| mlx-community-Qwen2.5-VL-7B-Instruct-8bit.jinja | Hermes 2 Pro |
+| mobiuslabsgmbh-DeepSeek-R1-ReDistill-Qwen-1.5B-v1.1.jinja | DeepSeek R1 (extract reasoning) |
+| netcat420-MFANNv0.20.jinja | Generic |
+| netcat420-MFANNv0.24.jinja | Generic |
+| netease-youdao-Confucius-o1-14B.jinja | Hermes 2 Pro |
+| nvidia-AceMath-7B-RM.jinja | Hermes 2 Pro |
+| nvidia-Eagle2-1B.jinja | Hermes 2 Pro |
+| nvidia-Eagle2-9B.jinja | Hermes 2 Pro |
+| nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | Llama 3.x |
+| onnx-community-DeepSeek-R1-Distill-Qwen-1.5B-ONNX.jinja | DeepSeek R1 (extract reasoning) |
+| open-thoughts-OpenThinker-7B.jinja | Hermes 2 Pro |
+| openchat-openchat-3.5-0106.jinja | Generic |
+| pankajmathur-orca_mini_v6_8b.jinja | Generic |
+| princeton-nlp-Mistral-7B-Base-SFT-RDPO.jinja | Generic |
+| princeton-nlp-Mistral-7B-Instruct-DPO.jinja | Generic |
+| princeton-nlp-Mistral-7B-Instruct-RDPO.jinja | Generic |
+| prithivMLmods-Bellatrix-Tiny-1.5B-R1.jinja | Hermes 2 Pro |
+| prithivMLmods-Bellatrix-Tiny-1B-R1.jinja | Llama 3.x |
+| prithivMLmods-Bellatrix-Tiny-1B-v3.jinja | Generic |
+| prithivMLmods-Bellatrix-Tiny-3B-R1.jinja | Llama 3.x |
+| prithivMLmods-Blaze-14B-xElite.jinja | Generic |
+| prithivMLmods-Calcium-Opus-14B-Elite2-R1.jinja | Hermes 2 Pro |
+| prithivMLmods-Calme-Ties-78B.jinja | Generic |
+| prithivMLmods-Calme-Ties2-78B.jinja | Generic |
+| prithivMLmods-Calme-Ties3-78B.jinja | Generic |
+| prithivMLmods-ChemQwen2-vL.jinja | Generic |
+| prithivMLmods-GWQ2b.jinja | Generic |
+| prithivMLmods-LatexMind-2B-Codec.jinja | Generic |
+| prithivMLmods-Llama-3.2-6B-AlgoCode.jinja | Llama 3.x |
+| prithivMLmods-Megatron-Opus-14B-Exp.jinja | Hermes 2 Pro |
+| prithivMLmods-Megatron-Opus-14B-Stock.jinja | Hermes 2 Pro |
+| prithivMLmods-Megatron-Opus-7B-Exp.jinja | Hermes 2 Pro |
+| prithivMLmods-Omni-Reasoner-Merged.jinja | Hermes 2 Pro |
+| prithivMLmods-Omni-Reasoner4-Merged.jinja | Hermes 2 Pro |
+| prithivMLmods-Primal-Opus-14B-Optimus-v1.jinja | Hermes 2 Pro |
+| prithivMLmods-QwQ-Math-IO-500M.jinja | Hermes 2 Pro |
+| prithivMLmods-Qwen-7B-Distill-Reasoner.jinja | DeepSeek R1 (extract reasoning) |
+| prithivMLmods-Qwen2.5-1.5B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
+| prithivMLmods-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
+| prithivMLmods-Qwen2.5-32B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
+| prithivMLmods-Qwen2.5-7B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
+| prithivMLmods-Triangulum-v2-10B.jinja | Hermes 2 Pro |
+| qingy2024-Falcon3-2x10B-MoE-Instruct.jinja | Hermes 2 Pro |
+| rubenroy-Zurich-14B-GCv2-5m.jinja | Hermes 2 Pro |
+| rubenroy-Zurich-7B-GCv2-5m.jinja | Hermes 2 Pro |
+| silma-ai-SILMA-Kashif-2B-Instruct-v1.0.jinja | Generic |
+| simplescaling-s1-32B.jinja | Hermes 2 Pro |
+| sometimesanotion-Lamarck-14B-v0.7.jinja | Hermes 2 Pro |
+| sonthenguyen-zephyr-sft-bnb-4bit-DPO-mtbr-180steps.jinja | Generic |
+| sthenno-tempesthenno-icy-0130.jinja | Generic |
+| sumink-qwft.jinja | Hermes 2 Pro |
+| teknium-OpenHermes-2.5-Mistral-7B.jinja | Generic |
+| thirdeyeai-elevate360m.jinja | Generic |
+| tiiuae-Falcon3-10B-Instruct.jinja | Hermes 2 Pro |
+| unsloth-DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit.jinja | DeepSeek R1 (extract reasoning) |
+| unsloth-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
+| unsloth-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
+| unsloth-Mistral-Small-24B-Instruct-2501-unsloth-bnb-4bit.jinja | Generic |
+| upstage-solar-pro-preview-instruct.jinja | Generic |
+| whyhow-ai-PatientSeek.jinja | Generic |
+| xwen-team-Xwen-72B-Chat.jinja | Hermes 2 Pro |
+| xwen-team-Xwen-7B-Chat.jinja | Hermes 2 Pro |
+
+This table can be generated with:
+
+```bash
+./build/bin/test-chat ../minja/build/tests/*.jinja 2>/dev/null
+```
+
+</details>
+
+# Usage - need tool-aware Jinja template
+
+First, start a server with any model, but make sure it has a tools-enabled template: you can verify this by inspecting the `chat_template` or `chat_template_tool_use` properties in `http://localhost:8080/props`).
+
+Here are some models known to work (w/ chat template override when needed):
+
+```shell
+# Native support:
+
+llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M
+llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L
+llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M
+llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M
+
+# Native support for DeepSeek R1 works best w/ our own template (official template buggy)
+
+llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \
+--chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
+
+llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \
+--chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
+
+# Native support requires the right template for these GGUFs:
+
+llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \
+--chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use )
+
+llama-server --jinja -fa -hf bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M \
+--chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use )
+
+llama-server --jinja -fa -hf bartowski/firefunction-v2-GGUF -hff firefunction-v2-IQ1_M.gguf \
+--chat-template-file <( python scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 tool_use )
+
+llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L \
+--chat-template-file <( python scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use )
+
+# Generic format support
+llama-server --jinja -fa -hf bartowski/phi-4-GGUF:Q4_0
+llama-server --jinja -fa -hf bartowski/gemma-2-2b-it-GGUF:Q8_0
+llama-server --jinja -fa -hf bartowski/c4ai-command-r-v01-GGUF:Q2_K
+```
+
+> [!TIP]
+> If there is no official `tool_use` Jinja template, you may want to set `--chat-template chatml` to use a default that works with many models (YMMV!), or write your own (e.g. we provide a custom [llama-cpp-deepseek-r1.jinja](../models/templates/llama-cpp-deepseek-r1.jinja) for DeepSeek R1 distills)
+
+Test in CLI (or with any library / software that can use OpenAI-compatible API backends):
+
+```bash
+curl http://localhost:8080/v1/chat/completions -d '{
+"model": "gpt-3.5-turbo",
+"tools": [
+    {
+    "type":"function",
+    "function":{
+        "name":"python",
+        "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
+        "parameters":{
+        "type":"object",
+        "properties":{
+            "code":{
+            "type":"string",
+            "description":"The code to run in the ipython interpreter."
+            }
+        },
+        "required":["code"]
+        }
+    }
+    }
+],
+"messages": [
+    {
+    "role": "user",
+    "content": "Print a hello world message with python."
+    }
+]
+}'
+```
+
+<details>
+<summary>Show output</summary>
+
+```json
+{
+"choices": [
+    {
+    "finish_reason": "tool",
+    "index": 0,
+    "message": {
+        "content": null,
+        "tool_calls": [
+        {
+            "name": "python",
+            "arguments": "{\"code\":\" \\nprint(\\\"Hello, World!\\\")\"}"
+        }
+        ],
+        "role": "assistant"
+    }
+    }
+],
+"created": 1727287211,
+"model": "gpt-3.5-turbo",
+"object": "chat.completion",
+"usage": {
+    "completion_tokens": 16,
+    "prompt_tokens": 44,
+    "total_tokens": 60
+},
+"id": "chatcmpl-Htbgh9feMmGM0LEH2hmQvwsCxq3c6Ni8"
+}
+```
+
+</details>
@@ -132,7 +132,7 @@ int main(int argc, char ** argv) {

                const auto t_pp_start = ggml_time_us();

-                llama_kv_cache_clear(ctx);
+                llama_kv_self_clear(ctx);

                if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                    LOG_ERR("%s: llama_decode() failed\n", __func__);
@@ -141,7 +141,7 @@ int main(int argc, char ** argv) {

                if (is_pp_shared) {
                    for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
                    }
                }

@@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 {
 }

 for i in 1 ..< n_parallel {
-    llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
 }

 if n_parallel > 1 {
@@ -342,7 +342,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 }

 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
        fprintf(stderr, "%s : failed to eval\n", __func__);
        return false;
@@ -37,7 +37,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
    const struct llama_model * model = llama_get_model(ctx);

    // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);

    // run model
    LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
@@ -45,7 +45,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
        }

        // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
        llama_set_embeddings(ctx, true);
        llama_set_causal_attn(ctx, false);

@@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std

    llama_token eos_token = llama_vocab_eos(vocab);

-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);
    llama_set_embeddings(ctx, false);
    llama_set_causal_attn(ctx, true);

@@ -498,7 +498,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
        const auto t_start = std::chrono::high_resolution_clock::now();

        // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        llama_batch batch = llama_batch_init(n_batch, 0, 1);

@@ -332,8 +332,8 @@ int main(int argc, char ** argv) {
                LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                    n_past, n_left, n_ctx, params.n_keep, n_discard);

-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);

                n_past -= n_discard;

@@ -1578,7 +1578,7 @@ int main(int argc, char ** argv) {

        test t(inst, lmodel, ctx);

-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        // cool off before the test
        if (params.delay) {
@@ -1618,7 +1618,7 @@ int main(int argc, char ** argv) {
        }

        for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_clear(ctx);
+            llama_kv_self_clear(ctx);

            uint64_t t_start = get_time_ns();

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
        }

        batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);

        const auto t_pp_start = ggml_time_us();
        if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(

        LOGi("Benchmark text generation (tg)");

-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);
        const auto t_tg_start = ggml_time_us();
        for (i = 0; i < tg; i++) {

@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(

        const auto t_tg_end = ggml_time_us();

-        llama_kv_cache_clear(context);
+        llama_kv_self_clear(context);

        const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
        const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
+    llama_kv_self_clear(reinterpret_cast<llama_context *>(context));
 }
@@ -210,7 +210,7 @@ actor LlamaContext {
            }
            batch.logits[Int(batch.n_tokens) - 1] = 1 // true

-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)

            let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;

@@ -223,7 +223,7 @@ actor LlamaContext {

            // bench text generation

-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)

            let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;

@@ -242,7 +242,7 @@ actor LlamaContext {

            let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;

-            llama_kv_cache_clear(context)
+            llama_kv_self_clear(context)

            let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
            let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -292,7 +292,7 @@ actor LlamaContext {
    func clear() {
        tokens_list.removeAll()
        temporary_invalid_cchars.removeAll()
-        llama_kv_cache_clear(context)
+        llama_kv_self_clear(context)
    }

    private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
@@ -0,0 +1,183 @@
+# Granite Vision
+
+Download the model and point your `GRANITE_MODEL` environment variable to the path.
+
+```bash
+$ git clone https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview
+$ export GRANITE_MODEL=./granite-vision-3.1-2b-preview
+```
+
+
+### 1. Running llava surgery v2.
+First, we need to run the llava surgery script as shown below:
+
+`python llava_surgery_v2.py -C -m $GRANITE_MODEL`
+
+You should see two new files (`llava.clip` and `llava.projector`) written into your model's directory, as shown below.
+
+```bash
+$ ls $GRANITE_MODEL | grep -i llava
+llava.clip
+llava.projector
+```
+
+We should see that the projector and visual encoder get split out into the llava files. Quick check to make sure they aren't empty:
+```python
+import os
+import torch
+
+MODEL_PATH = os.getenv("GRANITE_MODEL")
+if not MODEL_PATH:
+    raise ValueError("env var GRANITE_MODEL is unset!")
+
+encoder_tensors = torch.load(os.path.join(MODEL_PATH, "llava.clip"))
+projector_tensors = torch.load(os.path.join(MODEL_PATH, "llava.projector"))
+
+assert len(encoder_tensors) > 0
+assert len(projector_tensors) > 0
+```
+
+If you actually inspect the `.keys()` of the loaded tensors, you should see a lot of `vision_model` tensors in the `encoder_tensors`, and 5 tensors (`'multi_modal_projector.linear_1.bias'`, `'multi_modal_projector.linear_1.weight'`, `'multi_modal_projector.linear_2.bias'`, `'multi_modal_projector.linear_2.weight'`, `'image_newline'`) in the multimodal `projector_tensors`.
+
+
+### 2. Creating the Visual Component GGUF
+To create the GGUF for the visual components, we need to write a config for the visual encoder; make sure the config contains the correct `image_grid_pinpoints`
+
+
+Note: we refer to this file as `$VISION_CONFIG` later on.
+```json
+{
+    "_name_or_path": "siglip-model",
+    "architectures": [
+      "SiglipVisionModel"
+    ],
+    "image_grid_pinpoints": [
+        [384,768],
+        [384,1152],
+        [384,1536],
+        [384,1920],
+        [384,2304],
+        [384,2688],
+        [384,3072],
+        [384,3456],
+        [384,3840],
+        [768,384],
+        [768,768],
+        [768,1152],
+        [768,1536],
+        [768,1920],
+        [1152,384],
+        [1152,768],
+        [1152,1152],
+        [1536,384],
+        [1536,768],
+        [1920,384],
+        [1920,768],
+        [2304,384],
+        [2688,384],
+        [3072,384],
+        [3456,384],
+        [3840,384]
+    ],
+    "mm_patch_merge_type": "spatial_unpad",
+    "hidden_size": 1152,
+    "image_size": 384,
+    "intermediate_size": 4304,
+    "model_type": "siglip_vision_model",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "patch_size": 14,
+    "layer_norm_eps": 1e-6,
+    "hidden_act": "gelu_pytorch_tanh",
+    "projection_dim": 0,
+    "vision_feature_layer": [-24, -20, -12, -1]
+}
+```
+
+Create a new directory to hold the visual components, and copy the llava.clip/projector files, as well as the vision config into it.
+
+```bash
+$ ENCODER_PATH=$PWD/visual_encoder
+$ mkdir $ENCODER_PATH
+
+$ cp $GRANITE_MODEL/llava.clip $ENCODER_PATH/pytorch_model.bin
+$ cp $GRANITE_MODEL/llava.projector $ENCODER_PATH/
+$ cp $VISION_CONFIG $ENCODER_PATH/config.json
+```
+
+At which point you should have something like this:
+```bash
+$ ls $ENCODER_PATH
+config.json             llava.projector         pytorch_model.bin
+```
+
+Now convert the components to GGUF; Note that we also override the image mean/std dev to `[.5,.5,.5]` since we use the siglip visual encoder - in the transformers model, you can find these numbers in the [preprocessor_config.json](https://huggingface.co/ibm-granite/granite-vision-3.1-2b-preview/blob/main/preprocessor_config.json).
+```bash
+$ python convert_image_encoder_to_gguf.py \
+    -m $ENCODER_PATH \
+    --llava-projector $ENCODER_PATH/llava.projector \
+    --output-dir $ENCODER_PATH \
+    --clip-model-is-vision \
+    --clip-model-is-siglip \
+    --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5
+```
+
+this will create the first GGUF file at `$ENCODER_PATH/mmproj-model-f16.gguf`; we will refer to the abs path of this file as the `$VISUAL_GGUF_PATH.`
+
+
+### 3. Creating the LLM GGUF.
+The granite vision model contains a granite LLM as its language model. For now, the easiest way to get the GGUF for LLM is by loading the composite model in `transformers` and exporting the LLM so that it can be directly converted with the normal conversion path.
+
+First, set the `LLM_EXPORT_PATH` to the path to export the `transformers` LLM to.
+```
+$ export LLM_EXPORT_PATH=$PWD/granite_vision_llm
+```
+
+```python
+import os
+import transformers
+
+MODEL_PATH = os.getenv("GRANITE_MODEL")
+if not MODEL_PATH:
+    raise ValueError("env var GRANITE_MODEL is unset!")
+
+LLM_EXPORT_PATH = os.getenv("LLM_EXPORT_PATH")
+if not MODEL_PATH:
+    raise ValueError("env var LLM_EXPORT_PATH is unset!")
+
+tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_PATH)
+
+# NOTE: granite vision support was added to transformers very recently (4.49);
+# if you get size mismatches, your version is too old.
+# If you are running with an older version, set `ignore_mismatched_sizes=True`
+# as shown below; it won't be loaded correctly, but the LLM part of the model that
+# we are exporting will be loaded correctly.
+model = transformers.AutoModelForImageTextToText.from_pretrained(MODEL_PATH, ignore_mismatched_sizes=True)
+
+tokenizer.save_pretrained(LLM_EXPORT_PATH)
+model.language_model.save_pretrained(LLM_EXPORT_PATH)
+```
+
+Now you can convert the exported LLM to GGUF with the normal converter in the root of the llama cpp project.
+```bash
+$ LLM_GGUF_PATH=$LLM_EXPORT_PATH/granite_llm.gguf
+...
+$ python convert_hf_to_gguf.py --outfile $LLM_GGUF_PATH $LLM_EXPORT_PATH
+```
+
+
+### 4. Running the Model in Llama cpp
+Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. Sample usage:
+
+Note - the test image shown below can be found [here](https://github-production-user-asset-6210df.s3.amazonaws.com/10740300/415512792-d90d5562-8844-4f34-a0a5-77f62d5a58b5.jpg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20250221%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20250221T054145Z&X-Amz-Expires=300&X-Amz-Signature=86c60be490aa49ef7d53f25d6c973580a8273904fed11ed2453d0a38240ee40a&X-Amz-SignedHeaders=host).
+
+```bash
+$ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \
+    --mmproj $VISUAL_GGUF_PATH \
+    --image cherry_blossom.jpg \
+    -c 16384 \
+    -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\<image>\nWhat type of flowers are in this picture?\n<|assistant|>\n" \
+    --temp 0
+```
+
+Sample response: `The flowers in the picture are cherry blossoms, which are known for their delicate pink petals and are often associated with the beauty of spring.`
@@ -101,8 +101,27 @@ python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknow
 ```

 **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
+
 **note** llava-1.6 greatly benefits from batched prompt processing (defaults work)

+**note** if the language model in step `6)` is incompatible with the legacy conversion script, the easiest way handle the LLM model conversion is to load the model in transformers, and export only the LLM from the llava next model.
+
+```python
+import os
+import transformers
+
+model_path = ...
+llm_export_path = ...
+
+tokenizer = transformers.AutoTokenizer.from_pretrained(model_path)
+model = transformers.AutoModelForImageTextToText.from_pretrained(model_path)
+
+tokenizer.save_pretrained(llm_export_path)
+model.language_model.save_pretrained(llm_export_path)
+```
+
+Then, you can convert the LLM using the `convert_hf_to_gguf.py` script, which handles more LLM architectures.
+
 ## llava-cli templating and llava-1.6 prompting

 llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
@@ -40,6 +40,7 @@
 #include <map>
 #include <regex>
 #include <stdexcept>
+#include <unordered_set>
 #include <vector>
 #include <sstream>
 #include <cinttypes>
@@ -120,6 +121,7 @@ static std::string format(const char * fmt, ...) {
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"
 #define KEY_IMAGE_STD           "clip.vision.image_std"
 #define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"

 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
@@ -444,8 +446,9 @@ struct clip_hparams {

    char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default)

-    int32_t image_grid_pinpoints[32];
+    std::vector<int32_t> image_grid_pinpoints;
    int32_t image_crop_resolution;
+    std::unordered_set<int32_t> vision_feature_layer;
 };

 struct clip_layer {
@@ -585,6 +588,7 @@ struct clip_ctx {
    struct clip_vision_model vision_model;
    projector_type proj_type = PROJECTOR_TYPE_MLP;

+    int32_t max_feature_layer;
    float image_mean[3];
    float image_std[3];
    bool use_gelu = false;
@@ -651,7 +655,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
    const int hidden_size          = hparams.hidden_size;
    const int n_head               = hparams.n_head;
    const int d_head               = hidden_size / n_head;
-    int n_layer                    = hparams.n_layer;
    const float eps                = hparams.eps;
    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};

@@ -752,13 +755,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
    }

+    std::vector<struct ggml_tensor *> embedding_stack;
+    const auto & vision_feature_layer = hparams.vision_feature_layer;
+
    // loop over layers
-    if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) {
-        n_layer += 1;
-    }
-    for (int il = 0; il < n_layer - 1; il++) {
+    for (int il = 0; il < ctx->max_feature_layer; il++) {
        struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states

+        // If this is an embedding feature layer, save the output.
+        // NOTE: 0 index here refers to the input to the encoder.
+        if (vision_feature_layer.find(il) != vision_feature_layer.end()) {
+            embedding_stack.push_back(embeddings);
+        }
+
        //const size_t nb_q_w = model.layers[il].q_w->nb[0];

        // layernorm1
@@ -846,7 +855,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        cur = ggml_add(ctx0, embeddings, cur);

        embeddings = cur;
-
    }

    // post-layernorm
@@ -857,6 +865,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
    }

+    // final layer is a vision feature layer
+    if (vision_feature_layer.find(ctx->max_feature_layer) != vision_feature_layer.end()) {
+        embedding_stack.push_back(embeddings);
+    }
+
+    // If feature layers are explicitly set, stack them (if we have multiple)
+    if (!embedding_stack.empty()) {
+        embeddings = embedding_stack[0];
+        for (size_t i = 1; i < embedding_stack.size(); i++) {
+            embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0);
+        }
+    }
+
    // llava projector
    if (ctx->has_llava_projector) {
        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
@@ -1443,14 +1464,26 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
            int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS);
            int n = gguf_get_arr_n(ctx, idx);
            const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx);
-            for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) {
-                hparams.image_grid_pinpoints[i] = pinpoints[i];
+            for (int i = 0; i < n; ++i) {
+                hparams.image_grid_pinpoints.push_back(pinpoints[i]);
            }
-            if (n < 32)
-                hparams.image_grid_pinpoints[n] = 0;
-        } catch (std::runtime_error & /*e*/) {
-            hparams.image_grid_pinpoints[0]=0;
-        }
+        } catch (std::runtime_error & /*e*/) { }
+
+        // Load the vision feature layer indices if they are explicitly provided;
+        // if multiple vision feature layers are present, the values will be concatenated
+        // to form the final visual features.
+        // NOTE: gguf conversions should standardize the values of the vision feature layer to
+        // be non-negative, since we use -1 to mark values as unset here.
+        try {
+            int idx = get_key_idx(ctx, KEY_FEATURE_LAYER);
+            int n = gguf_get_arr_n(ctx, idx);
+
+            const int32_t * vision_feature_layer = (const int32_t *)gguf_get_arr_data(ctx, idx);
+
+            for (int i = 0; i < n; ++i) {
+                hparams.vision_feature_layer.insert(vision_feature_layer[i]);
+            }
+        } catch (std::runtime_error & /*e*/) { }

        try {
            int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE);
@@ -1476,6 +1509,9 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
            new_clip->image_std[i]  = std_data[i];
        }

+        // Calculate the deepest feature layer based on hparams and projector type
+        new_clip->max_feature_layer = get_deepest_feature_layer(new_clip);
+
        if (verbosity >= 2) {
            LOG_INF("\n%s: vision model hparams\n", __func__);
            LOG_INF("image_size         %d\n", hparams.image_size);
@@ -1489,8 +1525,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
            LOG_INF("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
            LOG_INF("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
            LOG_INF("v_image_grid_pinpoints: ");
-            for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
-                LOG_INF("%d ", hparams.image_grid_pinpoints[i]);
+            for (const auto & pp : hparams.image_grid_pinpoints) {
+                LOG_INF("%d ", pp);
+            }
+            LOG_INF("\n");
+            LOG_INF("v_vision_feature_layer: ");
+            for (const auto & feature_layer: hparams.vision_feature_layer) {
+                LOG_INF("%d ", feature_layer);
            }
            LOG_INF("\n");
            LOG_INF("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
@@ -1729,11 +1770,11 @@ void clip_image_f32_batch_free(struct clip_image_f32_batch  * batch) {
    }
 }

-static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
+void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, clip_image_u8 * img) {
    img->nx = nx;
    img->ny = ny;
    img->buf.resize(3 * nx * ny);
-    memcpy(img->buf.data(), data, img->buf.size());
+    memcpy(img->buf.data(), rgb_pixels, img->buf.size());
 }

 bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
@@ -1743,7 +1784,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
        LOG_ERR("%s: failed to load image '%s'\n", __func__, fname);
        return false;
    }
-    build_clip_img_from_data(data, nx, ny, img);
+    clip_build_img_from_pixels(data, nx, ny, img);
    stbi_image_free(data);
    return true;
 }
@@ -1755,7 +1796,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
        LOG_ERR("%s: failed to decode image bytes\n", __func__);
        return false;
    }
-    build_clip_img_from_data(data, nx, ny, img);
+    clip_build_img_from_pixels(data, nx, ny, img);
    stbi_image_free(data);
    return true;
 }
@@ -2235,10 +2276,10 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
            }
        }
    } else {
-        if (params.image_grid_pinpoints[0] != 0) {
+        if (!params.image_grid_pinpoints.empty()) {
            // "spatial_unpad" with "anyres" processing for llava-1.6
            std::vector<std::pair<int, int>> possible_resolutions;
-            for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            for (size_t i = 0; i < params.image_grid_pinpoints.size(); i+=2) {
                possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
            }
            std::pair<int, int> best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions);
@@ -2404,7 +2445,14 @@ const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
 }

 const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
-    return ctx->vision_model.hparams.image_grid_pinpoints;
+    if (ctx->vision_model.hparams.image_grid_pinpoints.size()) {
+        return &ctx->vision_model.hparams.image_grid_pinpoints.front();
+    }
+    return nullptr;
+}
+
+size_t get_clip_image_grid_size(const struct clip_ctx * ctx) {
+    return ctx->vision_model.hparams.image_grid_pinpoints.size();
 }

 int clip_n_patches(const struct clip_ctx * ctx) {
@@ -2929,6 +2977,28 @@ bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
    return ctx->has_qwen2vl_merger;
 }

+// Determine the number of encoder layers to iterate over
+int get_deepest_feature_layer(const struct clip_ctx * ctx) {
+    // Get the index of the second to last layer; this is the
+    // default for models that have a llava projector
+    const auto & hparams = ctx->vision_model.hparams;
+    int n_layer = hparams.n_layer - 1;
+    int deepest_feature_layer = -1;
+
+    // Handle other projectors; incrementing here indicates that we
+    // should use the last encoder layer for the vision features.
+    if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) {
+        n_layer += 1;
+    }
+
+    // If we set explicit vision feature layers, only go up to the deepest one
+    for (const auto & feature_layer : hparams.vision_feature_layer) {
+        if (feature_layer > deepest_feature_layer) {
+            deepest_feature_layer = feature_layer;
+        }
+    }
+    return deepest_feature_layer < 0 ? n_layer : deepest_feature_layer;
+}

 bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) {
    clip_image_f32 clip_img;
@@ -55,6 +55,7 @@ CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);

 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);

 CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
 CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
@@ -73,6 +74,12 @@ CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
 CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);

+/**
+ * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
+ * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
+ */
+CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
+
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);

 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
@@ -89,11 +96,13 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);

 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);

+CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
+
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);

-CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);

 #ifdef __cplusplus
 }
@@ -6,7 +6,7 @@ import re
 import torch
 import numpy as np
 from gguf import *
-from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel
+from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel, SiglipVisionModel

 TEXT = "clip.text"
 VISION = "clip.vision"
@@ -37,6 +37,18 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b


 def get_tensor_name(name: str) -> str:
+    # Standardize the transformers llava next keys for
+    # image newline / mm projector with the classes in haotian-liu LLaVA
+    if name == "image_newline":
+        return "model.image_newline"
+    if name.startswith("multi_modal_projector"):
+        name = name.replace("multi_modal_projector", "mm")
+        if "linear_1" in name:
+            name = name.replace("linear_1", "0")
+        if "linear_2" in name:
+            name = name.replace("linear_2", "2")
+        return name
+
    if "projection" in name:
        return name
    if "mm_projector" in name:
@@ -83,8 +95,14 @@ ap.add_argument("--vision-only", action="store_true", required=False,
                help="Save a vision-only model. It can't be used to encode texts")
 ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
-ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+
+# Selectable visual encoders that are compatible with this script
+encoder_group = ap.add_mutually_exclusive_group()
+encoder_group.add_argument("--clip-model-is-openclip", action="store_true", required=False,
                help="The clip model is from openclip (for ViT-SO400M type))")
+encoder_group.add_argument("--clip-model-is-siglip", action="store_true", required=False,
+                help="the visual encoder is Siglip.")
+
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
 ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
@@ -109,7 +127,12 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir

-if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+if (
+    args.clip_model_is_vision or
+    not os.path.exists(dir_model + "/vocab.json") or
+    args.clip_model_is_openclip or
+    args.clip_model_is_siglip
+):
    vocab = None
    tokens = None
 else:
@@ -137,7 +160,10 @@ ftype = 1
 if args.use_f32:
    ftype = 0

-if args.clip_model_is_vision or args.clip_model_is_openclip:
+if args.clip_model_is_siglip:
+    model = SiglipVisionModel.from_pretrained(dir_model)
+    processor = None
+elif args.clip_model_is_vision or args.clip_model_is_openclip:
    model = CLIPVisionModel.from_pretrained(dir_model)
    processor = None
 else:
@@ -187,26 +213,71 @@ else:
 if has_text_encoder:
    assert t_hparams is not None
    assert tokens is not None
+    if args.clip_model_is_siglip:
+        text_projection_dim = 0
+    else:
+        text_projection_dim = t_hparams.get("projection_dim", config["projection_dim"])
    # text_model hparams
    fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
-    fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32("clip.text.projection_dim", text_projection_dim)
    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
    fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
    fout.add_token_list(tokens)

+
+
+def get_non_negative_vision_feature_layers(v_hparams):
+    """
+    Determine the vision feature layer(s) for the llava model, which are indices into the
+    hidden states of the visual encoder. Note that the hidden states array generally takes the
+    form:
+
+        [<emb input>, <output of enc block 0>, ... <output of enc block num_hidden_layers>]
+
+    so feature indices should be offset as n+1 to get the output of encoder block n.
+    We convert all vision feature layers to non-negative so that -1 can be used in
+    the model as an unset value. If no vision feature layer is found, we leave it unset.
+    """
+    num_hidden_layers = v_hparams["num_hidden_layers"]
+    to_non_negative = lambda layer_idx: layer_idx  if layer_idx >= 0 else num_hidden_layers + layer_idx + 1
+    feature_layers_key = None
+    # Key used for llava models in transformers
+    if "vision_feature_layer" in config:
+        feature_layers_key = "vision_feature_layer"
+    # Key used for llava models in the original format
+    elif "mm_vision_select_layer" in config:
+        feature_layers_key = "mm_vision_select_layer"
+    if feature_layers_key is not None:
+        feature_layers = config[feature_layers_key]
+        if isinstance(feature_layers, int):
+            feature_layers = [feature_layers]
+        return [to_non_negative(feature_layer) for feature_layer in feature_layers]
+
+# Determine if we have explicitly specified vision feature layers in our config
+feature_layers = get_non_negative_vision_feature_layers(v_hparams)
+
 if has_vision_encoder:
-    # vision_model hparams
+    # Siglip does not have a visual projector; set projection dim to 0
+    if args.clip_model_is_siglip:
+        visual_projection_dim = 0
+    else:
+        visual_projection_dim = v_hparams.get("projection_dim", config["projection_dim"])
+
+    # set vision_model hparams
    fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
    fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
-    fout.add_uint32("clip.vision.projection_dim", v_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32("clip.vision.projection_dim", visual_projection_dim)
    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
-    block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
+    if feature_layers:
+        block_count = max(feature_layers)
+    else:
+        block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
                            #     /**
                            #      "image_grid_pinpoints": [
@@ -258,7 +329,8 @@ if has_vision_encoder:
        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
    if "mm_projector_type" in v_hparams:
        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
-
+    if feature_layers:
+        fout.add_array("clip.vision.feature_layer", feature_layers)

    if processor is not None:
        image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean  # pyright: ignore[reportAttributeAccessIssue]
@@ -274,7 +346,13 @@ fout.add_bool("clip.use_gelu", use_gelu)


 if has_llava_projector:
-    model.vision_model.encoder.layers.pop(-1)
+    # By default, we drop the last layer for llava projector
+    # models unless we have explicitly set vision feature layers
+    if feature_layers is None:
+        model.vision_model.encoder.layers.pop(-1)
+    else:
+        model.vision_model.encoder.layers = model.vision_model.encoder.layers[:max(feature_layers)]
+
    projector = torch.load(args.llava_projector)
    for name, data in projector.items():
        name = get_tensor_name(name)
@@ -353,9 +353,10 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);

        const int32_t * image_grid = clip_image_grid(ctx_clip);
+        const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);

        std::vector<std::pair<int, int>> grid_pinpoints;
-        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+        for (size_t i = 0; i < num_gridpoints; i += 2) {
            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
        }

@@ -405,7 +406,8 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }

 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    int num_max_patches = 6;
+    // Granite vision uses up to 10 patches + base patch
+    int num_max_patches = 11;
    if (clip_is_minicpmv(ctx_clip)) {
        num_max_patches = 10;
    }
@@ -33,6 +33,33 @@ def save_model(model, file_path, file_type):
    else:
        torch.save(model, file_path)

+# Helpers to match weight names from specific components or
+# determine if a saved shard contains that component
+def is_vision_tower(weight_name):
+    return (
+        weight_name.startswith("model.vision_tower") or
+        weight_name.startswith("vit.") or
+        weight_name.startswith("vision_tower")
+    )
+
+def is_newline(weight_name):
+    return (
+        weight_name.startswith("model.image_newline") or
+        weight_name.startswith("image_newline")
+    )
+
+def is_mm_projector(weight_name):
+    return (
+        weight_name.startswith("model.mm_projector") or
+        weight_name.startswith("vision_proj.") or
+        weight_name.startswith("multi_modal_projector")
+    )
+
+def newline_criteria(checkpoint):
+    return any(is_newline(k) for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(is_mm_projector(k) for k in checkpoint.keys())

 # Adapted function to clean vision tower from checkpoint
 def clean_vision_tower_from_checkpoint(checkpoint_path):
@@ -40,7 +67,7 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
    # file_type = 'pytorch'
    model_path = os.path.dirname(checkpoint_path)
    print(f"Searching for vision tower tensors in {checkpoint_path}")
-    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
+    clip_tensors = [k for k, v in checkpoint.items() if is_vision_tower(k)]

    if len(clip_tensors) > 0:
        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
@@ -84,12 +111,6 @@ def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):

    return newline_checkpoint_path, projector_checkpoint_path

-def newline_criteria(checkpoint):
-    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
-
-def proj_criteria(checkpoint):
-    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
-

 # Command-line interface setup
 ap = argparse.ArgumentParser()
@@ -123,14 +144,14 @@ first_checkpoint = None
 if newline_checkpoint_path is not None:
    print(f"Taking newline from {newline_checkpoint_path}")
    first_checkpoint, file_type = load_model(newline_checkpoint_path)
-    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if is_newline(k)]

 # Load the checkpoint
 mm_tensors = []
 last_checkpoint = None
 if projector_checkpoint_path is not None:
    last_checkpoint, file_type = load_model(projector_checkpoint_path)
-    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
+    mm_tensors = [k for k, v in last_checkpoint.items() if is_mm_projector(k)]

 if len(mm_tensors) == 0:
    if last_checkpoint is not None:
@@ -155,5 +176,5 @@ if len(projector) > 0:
    save_model(projector, f"{args.model}/llava.projector", 'pytorch')

 print("Done!")
-print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")
@@ -95,7 +95,7 @@ int main(int argc, char ** argv) {
    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));

    for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
    }

    const auto t_enc_end = ggml_time_us();
@@ -437,17 +437,17 @@ int main(int argc, char ** argv) {

        // KV cache management
        // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_cache_seq_rm(ctx, -1, n_past, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_past, -1);

        if (seq_id_best != 0) {
            // if a verification token matched, we keep the best sequence and remove the rest
            // this leads to some KV cache fragmentation
-            llama_kv_cache_seq_keep(ctx, seq_id_best);
-            llama_kv_cache_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_cache_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_kv_self_seq_keep(ctx, seq_id_best);
+            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
+            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);

            for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
            }
        }
    }
@@ -192,7 +192,7 @@ int main(int argc, char ** argv){

        // KV cache management
        // clean the cache of draft tokens that weren't accepted
-        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+        llama_kv_self_seq_rm(ctx, 0, n_past, -1);

        common_batch_clear(batch_tgt);
        common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
@@ -330,7 +330,7 @@ int main(int argc, char ** argv) {
        }

        // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
+        llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1);
    }

    LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
@@ -573,8 +573,8 @@ int main(int argc, char ** argv) {
                    LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                            n_past, n_left, n_ctx, params.n_keep, n_discard);

-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_self_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_self_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);

                    n_past -= n_discard;

@@ -597,9 +597,9 @@ int main(int argc, char ** argv) {
                    LOG_DBG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                    LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);

-                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_self_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_self_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_self_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);

                    n_past -= bd;

@@ -201,7 +201,7 @@ int main(int argc, char ** argv) {

        // assign the system KV cache to all parallel sequences
        for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
        }

        LOG_INF("\n");
@@ -233,9 +233,9 @@ int main(int argc, char ** argv) {
        if (batch.n_tokens == 0) {
            // all sequences have ended - clear the entire KV cache
            for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, -1, -1);
+                llama_kv_self_seq_rm(ctx, i, -1, -1);
                // but keep the system prompt
-                llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
            }

            LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -371,8 +371,8 @@ int main(int argc, char ** argv) {
                    }

                    // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
+                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);

                    const auto t_main_end = ggml_time_us();

@@ -132,11 +132,11 @@ int main(int argc, char ** argv) {
            const int ib = i/n_batch - 1;
            const int bd = n_batch_grp*(n_grp - 1);

-            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update  (ctx);
+            llama_kv_self_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_self_update  (ctx);

-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
        }

        common_batch_clear(batch);
@@ -166,12 +166,12 @@ int main(int argc, char ** argv) {

        LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);

-        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-      //llama_kv_cache_defrag (ctx);
-        llama_kv_cache_update (ctx);
+        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+      //llama_kv_self_defrag (ctx);
+        llama_kv_self_update (ctx);

-        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;

        common_batch_clear(batch);

@@ -197,12 +197,12 @@ int main(int argc, char ** argv) {
        if (n_discard > 0) {
            LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);

-            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-          //llama_kv_cache_defrag (ctx);
-            llama_kv_cache_update (ctx);
+            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+          //llama_kv_self_defrag (ctx);
+            llama_kv_self_update (ctx);

-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
        }
    }

@@ -361,7 +361,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
        const auto t_start = std::chrono::high_resolution_clock::now();

        // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        llama_batch batch = llama_batch_init(n_batch, 0, 1);

@@ -547,7 +547,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
        const auto t_start = std::chrono::high_resolution_clock::now();

        // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        for (int j = 0; j < num_batches; ++j) {
            const int batch_start = start + j * n_batch;
@@ -924,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
            return;
        }

-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        // decode all tasks [i0, i1)
        if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1203,7 +1203,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
            return;
        }

-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        // decode all tasks [i0, i1)
        if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1575,7 +1575,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
            return;
        }

-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        // decode all tasks [i0, i1)
        if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1765,7 +1765,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
        }

        // clear the KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);

        llama_batch batch = llama_batch_init(n_batch, 0, 1);

@@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke

 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
    // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
+    llama_kv_self_clear(ctx);

    // run model
    LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
@@ -323,25 +323,17 @@ class File {
        return 0;
    }

-    std::string read_all(const std::string & filename){
-        open(filename, "r");
-        lock();
-        if (!file) {
-            printe("Error opening file '%s': %s", filename.c_str(), strerror(errno));
-            return "";
-        }
-
+    std::string to_string() {
        fseek(file, 0, SEEK_END);
-        size_t size = ftell(file);
+        const size_t size = ftell(file);
        fseek(file, 0, SEEK_SET);
-
        std::string out;
        out.resize(size);
-        size_t read_size = fread(&out[0], 1, size, file);
+        const size_t read_size = fread(&out[0], 1, size, file);
        if (read_size != size) {
-            printe("Error reading file '%s': %s", filename.c_str(), strerror(errno));
-            return "";
+            printe("Error reading file: %s", strerror(errno));
        }
+
        return out;
    }

@@ -899,7 +891,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                           std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_get_kv_cache_used_cells(llama_data.context.get()) == 0;
+    const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0;

    const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
    prompt_tokens.resize(n_prompt_tokens);
@@ -915,7 +907,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt
 // Check if we have enough space in the context to evaluate this batch
 static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
    const int n_ctx      = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
+    const int n_ctx_used = llama_kv_self_used_cells(ctx.get());
    if (n_ctx_used + batch.n_tokens > n_ctx) {
        printf(LOG_COL_DEFAULT "\n");
        printe("context size exceeded\n");
@@ -985,7 +977,8 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str
 }

 static int read_user_input(std::string & user_input) {
-    static const char * prompt_prefix = "> ";
+    static const char * prompt_prefix_env = std::getenv("LLAMA_PROMPT_PREFIX");
+    static const char * prompt_prefix     = prompt_prefix_env ? prompt_prefix_env : "> ";
 #ifdef WIN32
    printf("\r" LOG_CLR_TO_EOL LOG_COL_DEFAULT "%s", prompt_prefix);

@@ -1098,59 +1091,66 @@ static int get_user_input(std::string & user_input, const std::string & user) {

 // Reads a chat template file to be used
 static std::string read_chat_template_file(const std::string & chat_template_file) {
-    if(chat_template_file.empty()){
-        return "";
-    }
-
    File file;
-    std::string chat_template = "";
-    chat_template = file.read_all(chat_template_file);
-    if(chat_template.empty()){
+    if (!file.open(chat_template_file, "r")) {
        printe("Error opening chat template file '%s': %s", chat_template_file.c_str(), strerror(errno));
        return "";
    }
-    return chat_template;
+
+    return file.to_string();
+}
+
+static int process_user_message(const Opt & opt, const std::string & user_input, LlamaData & llama_data,
+                                const common_chat_templates_ptr & chat_templates, int & prev_len,
+                                const bool stdout_a_terminal) {
+    add_message("user", opt.user.empty() ? user_input : opt.user, llama_data);
+    int new_len;
+    if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, true, new_len, opt.use_jinja) < 0) {
+        return 1;
+    }
+
+    std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len);
+    std::string response;
+    if (generate_response(llama_data, prompt, response, stdout_a_terminal)) {
+        return 1;
+    }
+
+    if (!opt.user.empty()) {
+        return 2;
+    }
+
+    add_message("assistant", response, llama_data);
+    if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, false, prev_len, opt.use_jinja) < 0) {
+        return 1;
+    }
+
+    return 0;
 }

 // Main chat loop function
-static int chat_loop(LlamaData & llama_data, const std::string & user, const std::string & chat_template_file, bool use_jinja) {
+static int chat_loop(LlamaData & llama_data, const Opt & opt) {
    int prev_len = 0;
    llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get()));
-
-    std::string chat_template = "";
-    if(!chat_template_file.empty()){
-        chat_template = read_chat_template_file(chat_template_file);
+    std::string chat_template;
+    if (!opt.chat_template_file.empty()) {
+        chat_template = read_chat_template_file(opt.chat_template_file);
    }
-    auto chat_templates = common_chat_templates_init(llama_data.model.get(), chat_template.empty() ? nullptr : chat_template);

+    common_chat_templates_ptr chat_templates    = common_chat_templates_init(llama_data.model.get(), chat_template);
    static const bool stdout_a_terminal = is_stdout_a_terminal();
    while (true) {
        // Get user input
        std::string user_input;
-        if (get_user_input(user_input, user) == 1) {
+        if (get_user_input(user_input, opt.user) == 1) {
            return 0;
        }

-        add_message("user", user.empty() ? user_input : user, llama_data);
-        int new_len;
-        if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, true, new_len, use_jinja) < 0) {
+        const int ret = process_user_message(opt, user_input, llama_data, chat_templates, prev_len, stdout_a_terminal);
+        if (ret == 1) {
            return 1;
-        }
-
-        std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len);
-        std::string response;
-        if (generate_response(llama_data, prompt, response, stdout_a_terminal)) {
-            return 1;
-        }
-
-        if (!user.empty()) {
+        } else if (ret == 2) {
            break;
        }
-
-        add_message("assistant", response, llama_data);
-        if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, false, prev_len, use_jinja) < 0) {
-            return 1;
-        }
    }

    return 0;
@@ -1208,7 +1208,7 @@ int main(int argc, const char ** argv) {
        return 1;
    }

-    if (chat_loop(llama_data, opt.user, opt.chat_template_file, opt.use_jinja)) {
+    if (chat_loop(llama_data, opt)) {
        return 1;
    }

@@ -15,7 +15,7 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    print_build_info();
+    common_init();

    if (params.n_predict < 0) {
        params.n_predict = 16;
@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
        fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);

        // erase whole kv
-        llama_kv_cache_clear(ctx3);
+        llama_kv_self_clear(ctx3);
        fprintf(stderr, "%s : kv cache cleared\n", __func__);

        // restore kv into seq 1
@@ -13,6 +13,7 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 * Multimodal (wip)
 * Monitoring endpoints
 * Schema-constrained JSON response format
+ * [Function calling](../../docs/function-calling.md) / tool use for ~any model

 The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggml-org/llama.cpp/issues/4216).

@@ -1120,381 +1121,9 @@ curl http://localhost:8080/v1/chat/completions \

 *Tool call support*

-[Function calling](https://platform.openai.com/docs/guides/function-calling) is supported for all models (see https://github.com/ggml-org/llama.cpp/pull/9639):
+[OpenAI-style function calling](https://platform.openai.com/docs/guides/function-calling) is supported with the `--jinja` flag (and may require a `--chat-template-file` override to get the right tool-use compatible Jinja template; worst case, `--chat-template chatml` may also work).

- Requires `--jinja` flag
- Native tool call formats supported:
-  - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2
-  - Functionary v3.1 / v3.2
-  - Hermes 2/3, Qwen 2.5
-  - Mistral Nemo
-  - Firefunction v2
-  - Command R7B
-  - DeepSeek R1 (WIP / seems reluctant to call any tools?)
-
-  <details>
-  <summary>Show some common templates and which format handler they use</summary>
-
-  | Template | Format |
-  |----------|--------|
-  | Almawave-Velvet-14B.jinja | Hermes 2 Pro |
-  | AtlaAI-Selene-1-Mini-Llama-3.1-8B.jinja | Llama 3.x |
-  | CohereForAI-aya-expanse-8b.jinja | Generic |
-  | CohereForAI-c4ai-command-r-plus-default.jinja | Generic |
-  | CohereForAI-c4ai-command-r-plus-rag.jinja | Generic |
-  | CohereForAI-c4ai-command-r-plus-tool_use.jinja | Generic |
-  | CohereForAI-c4ai-command-r7b-12-2024-default.jinja | Command R7B (extract reasoning) |
-  | CohereForAI-c4ai-command-r7b-12-2024-rag.jinja | Command R7B (extract reasoning) |
-  | CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja | Command R7B (extract reasoning) |
-  | CohereForAI-c4ai-command-r7b-12-2024.jinja | Generic |
-  | DavieLion-Llama-3.2-1B-SPIN-iter3.jinja | Generic |
-  | Delta-Vector-Rei-12B.jinja | Mistral Nemo |
-  | EpistemeAI-Mistral-Nemo-Instruct-12B-Philosophy-Math.jinja | Mistral Nemo |
-  | FlofloB-83k_continued_pretraining_Qwen2.5-0.5B-Instruct_Unsloth_merged_16bit.jinja | Hermes 2 Pro |
-  | FlofloB-test_continued_pretraining_Phi-3-mini-4k-instruct_Unsloth_merged_16bit.jinja | Generic |
-  | HelpingAI-HAI-SER.jinja | Generic |
-  | HuggingFaceTB-SmolLM2-1.7B-Instruct.jinja | Generic |
-  | HuggingFaceTB-SmolLM2-135M-Instruct.jinja | Generic |
-  | HuggingFaceTB-SmolLM2-360M-Instruct.jinja | Generic |
-  | INSAIT-Institute-BgGPT-Gemma-2-27B-IT-v1.0.jinja | Generic |
-  | Ihor-Text2Graph-R1-Qwen2.5-0.5b.jinja | Hermes 2 Pro |
-  | Infinigence-Megrez-3B-Instruct.jinja | Generic |
-  | Josephgflowers-TinyLlama_v1.1_math_code-world-test-1.jinja | Generic |
-  | LGAI-EXAONE-EXAONE-3.5-2.4B-Instruct.jinja | Generic |
-  | LGAI-EXAONE-EXAONE-3.5-7.8B-Instruct.jinja | Generic |
-  | LatitudeGames-Wayfarer-12B.jinja | Generic |
-  | Magpie-Align-Llama-3-8B-Magpie-Align-v0.1.jinja | Generic |
-  | Magpie-Align-Llama-3.1-8B-Magpie-Align-v0.1.jinja | Generic |
-  | MaziyarPanahi-calme-3.2-instruct-78b.jinja | Generic |
-  | MiniMaxAI-MiniMax-Text-01.jinja | Generic |
-  | MiniMaxAI-MiniMax-VL-01.jinja | Generic |
-  | NaniDAO-deepseek-r1-qwen-2.5-32B-ablated.jinja | DeepSeek R1 (extract reasoning) |
-  | NexaAIDev-Octopus-v2.jinja | Generic |
-  | NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | Generic |
-  | NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | Hermes 2 Pro |
-  | NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | Generic |
-  | NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | Hermes 2 Pro |
-  | NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | Generic |
-  | NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | Hermes 2 Pro |
-  | NovaSky-AI-Sky-T1-32B-Flash.jinja | Hermes 2 Pro |
-  | NovaSky-AI-Sky-T1-32B-Preview.jinja | Hermes 2 Pro |
-  | OnlyCheeini-greesychat-turbo.jinja | Generic |
-  | Orenguteng-Llama-3.1-8B-Lexi-Uncensored-V2.jinja | Llama 3.x |
-  | OrionStarAI-Orion-14B-Chat.jinja | Generic |
-  | PowerInfer-SmallThinker-3B-Preview.jinja | Generic |
-  | PrimeIntellect-INTELLECT-1-Instruct.jinja | Generic |
-  | Qwen-QVQ-72B-Preview.jinja | Generic |
-  | Qwen-QwQ-32B-Preview.jinja | Hermes 2 Pro |
-  | Qwen-Qwen1.5-7B-Chat.jinja | Generic |
-  | Qwen-Qwen2-7B-Instruct.jinja | Generic |
-  | Qwen-Qwen2-VL-72B-Instruct.jinja | Generic |
-  | Qwen-Qwen2-VL-7B-Instruct.jinja | Generic |
-  | Qwen-Qwen2.5-0.5B.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-1.5B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-14B-Instruct-1M.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-14B.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-32B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-32B.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-3B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-72B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-7B-Instruct-1M.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-7B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-7B.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-Coder-32B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-Coder-7B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-Math-1.5B.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-Math-7B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-VL-3B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-VL-72B-Instruct.jinja | Hermes 2 Pro |
-  | Qwen-Qwen2.5-VL-7B-Instruct.jinja | Hermes 2 Pro |
-  | RWKV-Red-Team-ARWKV-7B-Preview-0.1.jinja | Hermes 2 Pro |
-  | SakanaAI-TinySwallow-1.5B-Instruct.jinja | Hermes 2 Pro |
-  | SakanaAI-TinySwallow-1.5B.jinja | Hermes 2 Pro |
-  | Sao10K-70B-L3.3-Cirrus-x1.jinja | Llama 3.x |
-  | SentientAGI-Dobby-Mini-Leashed-Llama-3.1-8B.jinja | Llama 3.x |
-  | SentientAGI-Dobby-Mini-Unhinged-Llama-3.1-8B.jinja | Llama 3.x |
-  | Steelskull-L3.3-Damascus-R1.jinja | Llama 3.x |
-  | Steelskull-L3.3-MS-Nevoria-70b.jinja | Llama 3.x |
-  | Steelskull-L3.3-Nevoria-R1-70b.jinja | Llama 3.x |
-  | THUDM-glm-4-9b-chat.jinja | Generic |
-  | THUDM-glm-edge-1.5b-chat.jinja | Generic |
-  | Tarek07-Progenitor-V1.1-LLaMa-70B.jinja | Llama 3.x |
-  | TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | Generic |
-  | TinyLlama-TinyLlama-1.1B-Chat-v1.0.jinja | Generic |
-  | UCLA-AGI-Mistral7B-PairRM-SPPO-Iter3.jinja | Generic |
-  | ValiantLabs-Llama3.1-8B-Enigma.jinja | Llama 3.x |
-  | abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | Generic |
-  | ai21labs-AI21-Jamba-1.5-Large.jinja | Generic |
-  | allenai-Llama-3.1-Tulu-3-405B-SFT.jinja | Generic |
-  | allenai-Llama-3.1-Tulu-3-405B.jinja | Generic |
-  | allenai-Llama-3.1-Tulu-3-8B.jinja | Generic |
-  | arcee-ai-Virtuoso-Lite.jinja | Hermes 2 Pro |
-  | arcee-ai-Virtuoso-Medium-v2.jinja | Hermes 2 Pro |
-  | arcee-ai-Virtuoso-Small-v2.jinja | Hermes 2 Pro |
-  | avemio-GRAG-NEMO-12B-ORPO-HESSIAN-AI.jinja | Generic |
-  | bespokelabs-Bespoke-Stratos-7B.jinja | Hermes 2 Pro |
-  | bfuzzy1-acheron-m1a-llama.jinja | Generic |
-  | bofenghuang-vigogne-2-70b-chat.jinja | Generic |
-  | bytedance-research-UI-TARS-72B-DPO.jinja | Generic |
-  | bytedance-research-UI-TARS-7B-DPO.jinja | Generic |
-  | bytedance-research-UI-TARS-7B-SFT.jinja | Generic |
-  | carsenk-phi3.5_mini_exp_825_uncensored.jinja | Generic |
-  | cyberagent-DeepSeek-R1-Distill-Qwen-14B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
-  | cyberagent-DeepSeek-R1-Distill-Qwen-32B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
-  | databricks-dbrx-instruct.jinja | Generic |
-  | deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | Generic |
-  | deepseek-ai-DeepSeek-Coder-V2-Lite-Base.jinja | Generic |
-  | deepseek-ai-DeepSeek-Coder-V2-Lite-Instruct.jinja | Generic |
-  | deepseek-ai-DeepSeek-R1-Distill-Llama-70B.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-R1-Distill-Qwen-14B.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-R1-Zero.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-V2-Lite.jinja | Generic |
-  | deepseek-ai-DeepSeek-V2.5.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-DeepSeek-V3.jinja | DeepSeek R1 (extract reasoning) |
-  | deepseek-ai-deepseek-coder-33b-instruct.jinja | Generic |
-  | deepseek-ai-deepseek-coder-6.7b-instruct.jinja | Generic |
-  | deepseek-ai-deepseek-coder-7b-instruct-v1.5.jinja | Generic |
-  | deepseek-ai-deepseek-llm-67b-chat.jinja | Generic |
-  | deepseek-ai-deepseek-llm-7b-chat.jinja | Generic |
-  | dicta-il-dictalm2.0-instruct.jinja | Generic |
-  | ehristoforu-Falcon3-8B-Franken-Basestruct.jinja | Hermes 2 Pro |
-  | fireworks-ai-llama-3-firefunction-v2.jinja | FireFunction v2 |
-  | godlikehhd-alpaca_data_sampled_ifd_new_5200.jinja | Hermes 2 Pro |
-  | godlikehhd-alpaca_data_score_max_0.7_2600.jinja | Hermes 2 Pro |
-  | google-gemma-2-27b-it.jinja | Generic |
-  | google-gemma-2-2b-it.jinja | Generic |
-  | google-gemma-2-2b-jpn-it.jinja | Generic |
-  | google-gemma-7b-it.jinja | Generic |
-  | huihui-ai-DeepSeek-R1-Distill-Llama-70B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
-  | huihui-ai-DeepSeek-R1-Distill-Llama-8B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
-  | huihui-ai-DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
-  | huihui-ai-DeepSeek-R1-Distill-Qwen-32B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
-  | huihui-ai-DeepSeek-R1-Distill-Qwen-7B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
-  | huihui-ai-Qwen2.5-14B-Instruct-1M-abliterated.jinja | Hermes 2 Pro |
-  | ibm-granite-granite-3.1-8b-instruct.jinja | Generic |
-  | indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | Generic |
-  | inflatebot-MN-12B-Mag-Mell-R1.jinja | Generic |
-  | jinaai-ReaderLM-v2.jinja | Generic |
-  | kms7530-chemeng_qwen-math-7b_24_1_100_1_nonmath.jinja | Hermes 2 Pro |
-  | knifeayumu-Cydonia-v1.3-Magnum-v4-22B.jinja | Mistral Nemo |
-  | langgptai-qwen1.5-7b-chat-sa-v0.1.jinja | Generic |
-  | lightblue-DeepSeek-R1-Distill-Qwen-7B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
-  | mattshumer-Reflection-Llama-3.1-70B.jinja | Generic |
-  | meetkai-functionary-medium-v3.1.jinja | Functionary v3.1 Llama 3.1 |
-  | meetkai-functionary-medium-v3.2.jinja | Functionary v3.2 |
-  | meta-llama-Llama-2-7b-chat-hf.jinja | Generic |
-  | meta-llama-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
-  | meta-llama-Llama-3.2-11B-Vision-Instruct.jinja | Llama 3.x |
-  | meta-llama-Llama-3.2-1B-Instruct.jinja | Llama 3.x |
-  | meta-llama-Llama-3.2-3B-Instruct.jinja | Llama 3.x |
-  | meta-llama-Llama-3.3-70B-Instruct.jinja | Llama 3.x |
-  | meta-llama-Meta-Llama-3-8B-Instruct.jinja | Generic |
-  | meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
-  | microsoft-Phi-3-medium-4k-instruct.jinja | Generic |
-  | microsoft-Phi-3-mini-4k-instruct.jinja | Generic |
-  | microsoft-Phi-3-small-8k-instruct.jinja | Generic |
-  | microsoft-Phi-3.5-mini-instruct.jinja | Generic |
-  | microsoft-Phi-3.5-vision-instruct.jinja | Generic |
-  | microsoft-phi-4.jinja | Generic |
-  | migtissera-Tess-3-Mistral-Nemo-12B.jinja | Generic |
-  | ministral-Ministral-3b-instruct.jinja | Generic |
-  | mistralai-Codestral-22B-v0.1.jinja | Generic |
-  | mistralai-Mistral-7B-Instruct-v0.1.jinja | Generic |
-  | mistralai-Mistral-7B-Instruct-v0.2.jinja | Generic |
-  | mistralai-Mistral-7B-Instruct-v0.3.jinja | Mistral Nemo |
-  | mistralai-Mistral-Large-Instruct-2407.jinja | Mistral Nemo |
-  | mistralai-Mistral-Large-Instruct-2411.jinja | Generic |
-  | mistralai-Mistral-Nemo-Instruct-2407.jinja | Mistral Nemo |
-  | mistralai-Mistral-Small-24B-Instruct-2501.jinja | Generic |
-  | mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | Generic |
-  | mkurman-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
-  | mlabonne-AlphaMonarch-7B.jinja | Generic |
-  | mlx-community-Josiefied-Qwen2.5-0.5B-Instruct-abliterated-v1-float32.jinja | Hermes 2 Pro |
-  | mlx-community-Qwen2.5-VL-7B-Instruct-8bit.jinja | Hermes 2 Pro |
-  | mobiuslabsgmbh-DeepSeek-R1-ReDistill-Qwen-1.5B-v1.1.jinja | DeepSeek R1 (extract reasoning) |
-  | netcat420-MFANNv0.20.jinja | Generic |
-  | netcat420-MFANNv0.24.jinja | Generic |
-  | netease-youdao-Confucius-o1-14B.jinja | Hermes 2 Pro |
-  | nvidia-AceMath-7B-RM.jinja | Hermes 2 Pro |
-  | nvidia-Eagle2-1B.jinja | Hermes 2 Pro |
-  | nvidia-Eagle2-9B.jinja | Hermes 2 Pro |
-  | nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | Llama 3.x |
-  | onnx-community-DeepSeek-R1-Distill-Qwen-1.5B-ONNX.jinja | DeepSeek R1 (extract reasoning) |
-  | open-thoughts-OpenThinker-7B.jinja | Hermes 2 Pro |
-  | openchat-openchat-3.5-0106.jinja | Generic |
-  | pankajmathur-orca_mini_v6_8b.jinja | Generic |
-  | princeton-nlp-Mistral-7B-Base-SFT-RDPO.jinja | Generic |
-  | princeton-nlp-Mistral-7B-Instruct-DPO.jinja | Generic |
-  | princeton-nlp-Mistral-7B-Instruct-RDPO.jinja | Generic |
-  | prithivMLmods-Bellatrix-Tiny-1.5B-R1.jinja | Hermes 2 Pro |
-  | prithivMLmods-Bellatrix-Tiny-1B-R1.jinja | Llama 3.x |
-  | prithivMLmods-Bellatrix-Tiny-1B-v3.jinja | Generic |
-  | prithivMLmods-Bellatrix-Tiny-3B-R1.jinja | Llama 3.x |
-  | prithivMLmods-Blaze-14B-xElite.jinja | Generic |
-  | prithivMLmods-Calcium-Opus-14B-Elite2-R1.jinja | Hermes 2 Pro |
-  | prithivMLmods-Calme-Ties-78B.jinja | Generic |
-  | prithivMLmods-Calme-Ties2-78B.jinja | Generic |
-  | prithivMLmods-Calme-Ties3-78B.jinja | Generic |
-  | prithivMLmods-ChemQwen2-vL.jinja | Generic |
-  | prithivMLmods-GWQ2b.jinja | Generic |
-  | prithivMLmods-LatexMind-2B-Codec.jinja | Generic |
-  | prithivMLmods-Llama-3.2-6B-AlgoCode.jinja | Llama 3.x |
-  | prithivMLmods-Megatron-Opus-14B-Exp.jinja | Hermes 2 Pro |
-  | prithivMLmods-Megatron-Opus-14B-Stock.jinja | Hermes 2 Pro |
-  | prithivMLmods-Megatron-Opus-7B-Exp.jinja | Hermes 2 Pro |
-  | prithivMLmods-Omni-Reasoner-Merged.jinja | Hermes 2 Pro |
-  | prithivMLmods-Omni-Reasoner4-Merged.jinja | Hermes 2 Pro |
-  | prithivMLmods-Primal-Opus-14B-Optimus-v1.jinja | Hermes 2 Pro |
-  | prithivMLmods-QwQ-Math-IO-500M.jinja | Hermes 2 Pro |
-  | prithivMLmods-Qwen-7B-Distill-Reasoner.jinja | DeepSeek R1 (extract reasoning) |
-  | prithivMLmods-Qwen2.5-1.5B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
-  | prithivMLmods-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
-  | prithivMLmods-Qwen2.5-32B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
-  | prithivMLmods-Qwen2.5-7B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
-  | prithivMLmods-Triangulum-v2-10B.jinja | Hermes 2 Pro |
-  | qingy2024-Falcon3-2x10B-MoE-Instruct.jinja | Hermes 2 Pro |
-  | rubenroy-Zurich-14B-GCv2-5m.jinja | Hermes 2 Pro |
-  | rubenroy-Zurich-7B-GCv2-5m.jinja | Hermes 2 Pro |
-  | silma-ai-SILMA-Kashif-2B-Instruct-v1.0.jinja | Generic |
-  | simplescaling-s1-32B.jinja | Hermes 2 Pro |
-  | sometimesanotion-Lamarck-14B-v0.7.jinja | Hermes 2 Pro |
-  | sonthenguyen-zephyr-sft-bnb-4bit-DPO-mtbr-180steps.jinja | Generic |
-  | sthenno-tempesthenno-icy-0130.jinja | Generic |
-  | sumink-qwft.jinja | Hermes 2 Pro |
-  | teknium-OpenHermes-2.5-Mistral-7B.jinja | Generic |
-  | thirdeyeai-elevate360m.jinja | Generic |
-  | tiiuae-Falcon3-10B-Instruct.jinja | Hermes 2 Pro |
-  | unsloth-DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit.jinja | DeepSeek R1 (extract reasoning) |
-  | unsloth-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
-  | unsloth-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
-  | unsloth-Mistral-Small-24B-Instruct-2501-unsloth-bnb-4bit.jinja | Generic |
-  | upstage-solar-pro-preview-instruct.jinja | Generic |
-  | whyhow-ai-PatientSeek.jinja | Generic |
-  | xwen-team-Xwen-72B-Chat.jinja | Hermes 2 Pro |
-  | xwen-team-Xwen-7B-Chat.jinja | Hermes 2 Pro |
-
-  This table can be generated with:
-
-  ```bash
-  ./build/bin/test-chat ../minja/build/tests/*.jinja 2>/dev/null
-  ```
-
-  </details>
-
- Generic tool call is supported when the template isn't recognized by native format handlers (you'll see `Chat format: Generic` in the logs).
-  - Use `--chat-template-file` to override the template when appropriate (see examples below)
-  - Generic support may consume more tokens and be less efficient than a model's native format.
-
- Run with:
-
-  ```shell
-  # Native support:
-
-  llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M
-  llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L
-  llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M
-  llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M
-
-  # Native support for DeepSeek R1 works best w/ our own template (official template buggy)
-
-  llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \
-    --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
-
-  llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \
-    --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
-
-  # Native support requires the right template for these GGUFs:
-
-  llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \
-    --chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use )
-
-  llama-server --jinja -fa -hf bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M \
-    --chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use )
-
-  llama-server --jinja -fa -hf bartowski/firefunction-v2-GGUF -hff firefunction-v2-IQ1_M.gguf \
-    --chat-template-file <( python scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 tool_use )
-
-  llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L \
-    --chat-template-file <( python scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use )
-
-  # Generic format support
-  llama-server --jinja -fa -hf bartowski/phi-4-GGUF:Q4_0
-  llama-server --jinja -fa -hf bartowski/gemma-2-2b-it-GGUF:Q8_0
-  llama-server --jinja -fa -hf bartowski/c4ai-command-r-v01-GGUF:Q2_K
-  ```
-
- Test in CLI:
-
-  ```bash
-  curl http://localhost:8080/v1/chat/completions -d '{
-    "model": "gpt-3.5-turbo",
-    "tools": [
-      {
-        "type":"function",
-        "function":{
-          "name":"python",
-          "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
-          "parameters":{
-            "type":"object",
-            "properties":{
-              "code":{
-                "type":"string",
-                "description":"The code to run in the ipython interpreter."
-              }
-            },
-            "required":["code"]
-          }
-        }
-      }
-    ],
-    "messages": [
-      {
-        "role": "user",
-        "content": "Print a hello world message with python."
-      }
-    ]
-  }'
-  ```
-
-  <details>
-  <summary>Show output</summary>
-
-  ```json
-  {
-    "choices": [
-      {
-        "finish_reason": "tool",
-        "index": 0,
-        "message": {
-          "content": null,
-          "tool_calls": [
-            {
-              "name": "python",
-              "arguments": "{\"code\":\" \\nprint(\\\"Hello, World!\\\")\"}"
-            }
-          ],
-          "role": "assistant"
-        }
-      }
-    ],
-    "created": 1727287211,
-    "model": "gpt-3.5-turbo",
-    "object": "chat.completion",
-    "usage": {
-      "completion_tokens": 16,
-      "prompt_tokens": 44,
-      "total_tokens": 60
-    },
-    "id": "chatcmpl-Htbgh9feMmGM0LEH2hmQvwsCxq3c6Ni8"
-  }
-  ```
-
-  </details>
+**See our [Function calling](../../docs/function-calling.md) docs** for more details, supported native tool call styles (generic tool call style is used as fallback) / examples of use.

 ### POST `/v1/embeddings`: OpenAI-compatible embeddings API

@@ -2083,7 +2083,7 @@ struct server_context {
        SRV_DBG("%s", "clearing KV cache\n");

        // clear the entire KV cache
-        llama_kv_cache_clear(ctx);
+        llama_kv_self_clear(ctx);
        clean_kv_cache = false;
    }

@@ -2625,8 +2625,8 @@ struct server_context {
                    res->n_tasks_deferred    = queue_tasks.queue_tasks_deferred.size();
                    res->t_start             = metrics.t_start;

-                    res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx);
-                    res->kv_cache_used_cells   = llama_get_kv_cache_used_cells(ctx);
+                    res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx);
+                    res->kv_cache_used_cells   = llama_kv_self_used_cells(ctx);

                    res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total;
                    res->t_prompt_processing_total       = metrics.t_prompt_processing_total;
@@ -2742,7 +2742,7 @@ struct server_context {

                    // Erase token cache
                    const size_t n_erased = slot->cache_tokens.size();
-                    llama_kv_cache_seq_rm(ctx, slot->id, -1, -1);
+                    llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
                    slot->cache_tokens.clear();

                    auto res = std::make_unique<server_task_result_slot_erase>();
@@ -2810,8 +2810,8 @@ struct server_context {

                SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);

-                llama_kv_cache_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
-                llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
+                llama_kv_self_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
+                llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);

                if (slot.params.cache_prompt) {
                    for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
@@ -3002,8 +3002,8 @@ struct server_context {

                                            const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;

-                                            llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c);
-                                            llama_kv_cache_seq_add(ctx, slot.id, head_c, -1,     kv_shift);
+                                            llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c);
+                                            llama_kv_self_seq_add(ctx, slot.id, head_c, -1,     kv_shift);

                                            for (size_t i = 0; i < n_match; i++) {
                                                slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
@@ -3041,9 +3041,9 @@ struct server_context {
                    }

                    // keep only the common part
-                    if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
+                    if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
                        // could not partially delete (likely using a non-Transformer model)
-                        llama_kv_cache_seq_rm(ctx, slot.id, -1, -1);
+                        llama_kv_self_seq_rm(ctx, slot.id, -1, -1);

                        // there is no common part left
                        slot.n_past = 0;
@@ -3283,7 +3283,7 @@ struct server_context {
                slot.cache_tokens.push_back(id);
                slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);

-                llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+                llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);

                for (size_t i = 0; i < ids.size(); ++i) {
                    completion_token_output result;
@@ -283,7 +283,7 @@ class ServerPreset:
        server.model_hf_repo = "ggml-org/models"
        server.model_hf_file = "tinyllamas/stories260K.gguf"
        server.model_alias = "tinyllama-2"
-        server.n_ctx = 256
+        server.n_ctx = 512
        server.n_batch = 32
        server.n_slots = 2
        server.n_predict = 64
@@ -7,6 +7,8 @@

 // increase max payload length to allow use of larger context size
 #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
+// disable Nagle's algorithm
+#define CPPHTTPLIB_TCP_NODELAY true
 #include "httplib.h"

 // Change JSON_ASSERT from assert() to GGML_ASSERT:
@@ -519,8 +521,13 @@ static json oaicompat_completion_params_parse(const json & body) {
        throw std::runtime_error("Only one completion choice is allowed");
    }

+    // Handle "echo" field
+    if (json_value(body, "echo", false)) {
+        throw std::runtime_error("Only no echo is supported");
+    }
+
    // Params supported by OAI but unsupported by llama.cpp
-    static const std::vector<std::string> unsupported_params { "best_of", "echo", "suffix" };
+    static const std::vector<std::string> unsupported_params { "best_of", "suffix" };
    for (const auto & param : unsupported_params) {
        if (body.contains(param)) {
            throw std::runtime_error("Unsupported param: " + param);
@@ -596,7 +603,7 @@ static json oaicompat_completion_params_parse(
    inputs.tool_choice           = common_chat_tool_choice_parse_oaicompat(json_value(body, "tool_choice", std::string("auto")));
    inputs.json_schema           = json_schema.is_null() ? "" : json_schema.dump();
    inputs.grammar               = grammar;
-    inputs.add_generation_prompt = true;
+    inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true);
    inputs.use_jinja             = use_jinja;
    inputs.parallel_tool_calls   = json_value(body, "parallel_tool_calls", false);
    inputs.extract_reasoning     = reasoning_format != COMMON_REASONING_FORMAT_NONE;
@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
    auto generate = [&](const std::string & prompt) {
        std::string response;

-        const bool is_first = llama_get_kv_cache_used_cells(ctx) == 0;
+        const bool is_first = llama_kv_self_used_cells(ctx) == 0;

        // tokenize the prompt
        const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,7 +113,7 @@ int main(int argc, char ** argv) {
        while (true) {
            // check if we have enough space in the context to evaluate this batch
            int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_get_kv_cache_used_cells(ctx);
+            int n_ctx_used = llama_kv_self_used_cells(ctx);
            if (n_ctx_used + batch.n_tokens > n_ctx) {
                printf("\033[0m\n");
                fprintf(stderr, "context size exceeded\n");
@@ -217,7 +217,7 @@ int main(int argc, char ** argv) {
        {
            LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);

-            llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
+            llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1);
        }

        if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
@@ -420,14 +420,14 @@ int main(int argc, char ** argv) {
            {
                LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);

-                llama_kv_cache_seq_keep(ctx_dft, s_keep);
-                llama_kv_cache_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
-                llama_kv_cache_seq_keep(ctx_dft, 0);
+                llama_kv_self_seq_keep(ctx_dft, s_keep);
+                llama_kv_self_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
+                llama_kv_self_seq_keep(ctx_dft, 0);

-                llama_kv_cache_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
-                llama_kv_cache_seq_keep(ctx_tgt, s_keep);
-                llama_kv_cache_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
-                llama_kv_cache_seq_keep(ctx_tgt, 0);
+                llama_kv_self_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
+                llama_kv_self_seq_keep(ctx_tgt, s_keep);
+                llama_kv_self_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
+                llama_kv_self_seq_keep(ctx_tgt, 0);
            }

            for (int s = 0; s < n_seq_dft; ++s) {
@@ -444,7 +444,7 @@ int main(int argc, char ** argv) {
            common_batch_clear(batch_dft);
            common_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);

-            llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
            // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
            llama_decode(ctx_dft, batch_dft);

@@ -503,8 +503,8 @@ int main(int argc, char ** argv) {
                    if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                        LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);

-                        llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+                        llama_kv_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
+                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);

                        // all previous tokens from this branch are now also part of the new branch
                        for (int t = 0; t < batch_tgt.n_tokens; ++t) {
@@ -585,9 +585,9 @@ int main(int argc, char ** argv) {

        // evaluate the target model on the drafted tokens
        {
-            llama_kv_cache_seq_keep(ctx_tgt, 0);
+            llama_kv_self_seq_keep(ctx_tgt, 0);
            for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
+                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
            }

            // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
@@ -3,7 +3,7 @@
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT
-
+export ONEAPI_DEVICE_SELECTOR="level_zero:0"
 source /opt/intel/oneapi/setvars.sh

 #export GGML_SYCL_DEBUG=1
@@ -13,7 +13,7 @@ source /opt/intel/oneapi/setvars.sh
 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
 MODEL_FILE=models/llama-2-7b.Q4_0.gguf
 NGL=33
-CONEXT=8192
+CONEXT=4096

 if [ $# -gt 0 ]; then
    GGML_SYCL_DEVICE=$1
@@ -122,6 +122,7 @@ endif()
 option(GGML_LASX             "ggml: enable lasx"             ON)
 option(GGML_LSX              "ggml: enable lsx"              ON)
 option(GGML_RVV              "ggml: enable rvv"              ON)
+option(GGML_VXE              "ggml: enable vxe"              ON)

 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
@@ -151,6 +152,7 @@ set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                            "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
 option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
+option(GGML_CUDA_FA                         "ggml: compile ggml FlashAttention CUDA kernels"  ON)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})

@@ -210,6 +212,8 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)

 find_package(Threads REQUIRED)

+include(GNUInstallDirs)
+
 #
 # build the library
 #
@@ -233,7 +237,6 @@ endif ()
 # install
 #

-include(GNUInstallDirs)
 include(CMakePackageConfigHelpers)

 # all public headers
@@ -112,7 +112,7 @@ foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})

    string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}")
    if(is_cpu_variant)
-        list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml" "ggml::ggml-base")
+        list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
        set_target_properties(ggml::${_ggml_backend}
           PROPERTIES
               INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}")
@@ -124,7 +124,7 @@ foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
        endif()

    else()
-        list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml" "ggml::ggml-base")
+        list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base")
        set_target_properties(ggml::${_ggml_backend}
            PROPERTIES
                INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}")
@@ -139,6 +139,11 @@ foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS})
    list(APPEND _ggml_all_targets ggml::${_ggml_backend})
 endforeach()

+list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}")
+set_target_properties(ggml::ggml
+    PROPERTIES
+        INTERFACE_LINK_LIBRARIES "${GGML_INTERFACE_LINK_LIBRARIES}")
+
 add_library(ggml::all INTERFACE IMPORTED)
 set_target_properties(ggml::all
    PROPERTIES
@@ -99,6 +99,7 @@ extern "C" {
    // other
    GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
    GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
    GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
    GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);

@@ -310,6 +310,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        if (GGML_RVV)
            list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
        endif()
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
+        message(STATUS "s390x detected")
+        file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
+        string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
+
+        # TODO: Separation to determine activation of VX/VXE/VXE2
+        if (${S390X_M} MATCHES "8561|8562")
+            message(STATUS "z15 target")
+            list(APPEND ARCH_FLAGS -march=z15 -mtune=z15)
+        elseif (${S390X_M} MATCHES "3931")
+            message(STATUS "z16 target")
+            list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
+        else()
+            message(STATUS "Unknown target")
+            message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
+            list(APPEND ARCH_FLAGS -march=native -mtune=native)
+        endif()
+
+        if (GGML_VXE)
+            list(APPEND ARCH_FLAGS -mvx -mzvector)
+        endif()
    else()
        message(STATUS "Unknown architecture")
    endif()
@@ -59,6 +59,15 @@ struct ggml_compute_params {
 #endif
 #endif

+#if defined(__s390x__) && defined(__VEC__)
+#ifndef __VXE__
+#define __VXE__
+#endif
+#ifndef __VXE2__
+#define __VXE2__
+#endif
+#endif
+
 #if defined(__ARM_FEATURE_SVE)
 #include <arm_sve.h>
 #include <sys/prctl.h>
@@ -359,6 +368,148 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #endif
 #endif

+#if defined(__VXE__) || defined(__VXE2__)
+#include <vecintrin.h>
+
+#define vec_neg(a)    (-(a))                // Vector Negate
+#define vec_add(a, b) ((a) + (b))           // Vector Add
+#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
+#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
+#define vec_div(a, b) ((a) / (b))           // Vector Divide
+#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
+#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
+#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
+#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
+#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
+
+#ifndef vec_and
+#define vec_and(a, b) ((a) & (b)) // Vector AND
+#endif
+
+#ifndef vec_or
+#define vec_or(a, b)  ((a) | (b)) // Vector OR
+#endif
+
+#ifndef vec_xor
+#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
+#endif
+
+typedef signed char char8x16_t __attribute__((vector_size(16)));
+typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
+
+typedef int8_t  int8x16_t __attribute__((vector_size(16)));
+typedef int16_t int16x8_t __attribute__((vector_size(16)));
+typedef int32_t int32x4_t __attribute__((vector_size(16)));
+
+typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
+typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
+typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
+
+typedef float float32x4_t __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute((vector_size(16)));
+
+typedef signed long long long64x2_t __attribute((vector_size(16)));
+typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+    res.val[2] = vec_xl(32, ptr);
+    res.val[3] = vec_xl(48, ptr);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+    res.val[2] = vec_xl(32, ptr);
+    res.val[3] = vec_xl(48, ptr);
+
+    return res;
+}
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vec_xl( 0, ptr);
+    res.val[1] = vec_xl(16, ptr);
+
+    return res;
+}
+
+/*
+    ! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
+    !          or iq4_nl for example implementation.
+*/
+inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
+    const uchar8x16_t v_maske = {  0,  1,  4,  5,  8,  9, 12, 13,
+                                  16, 17, 20, 21, 24, 25, 28, 29 };
+
+    const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
+    const int16x8_t v_abe = vec_perm(a, b, v_maske);
+    return v_abo + v_abe;
+}
+
+inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
+    return acc + (vec_unpackh(p) + vec_unpackl(p));
+}
+
+#endif
+
 #if defined(__loongarch_asx)
 /* float type data load instructions */
 static __m128 __lsx_vreplfr2vr_s(const float val) {
@@ -1011,6 +1011,38 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);

    }
+#elif defined(__VXE__) || defined(__VXE2__)
+    for (int i = 0; i < nb; i++) {
+        __vector float srcv [8];
+        __vector float asrcv[8];
+        __vector float amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f / d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const __vector float v = vec_mul(srcv[j], vec_splats(id));
+            const __vector int32_t vi = vec_signed(v);
+
+            y[i].qs[4*j + 0] = vec_extract(vi, 0);
+            y[i].qs[4*j + 1] = vec_extract(vi, 1);
+            y[i].qs[4*j + 2] = vec_extract(vi, 2);
+            y[i].qs[4*j + 3] = vec_extract(vi, 3);
+        }
+    }
 #else
    GGML_UNUSED(nb);
    // scalar
@@ -1337,6 +1369,44 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
    }
+#elif defined(__VXE__) || defined(__VXE2__)
+    for (int i = 0; i < nb; i++) {
+        __vector float srcv [8];
+        __vector float asrcv[8];
+        __vector float amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f / d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        __vector int32_t acc = vec_splats(0);
+
+        for (int j = 0; j < 8; j++) {
+            const __vector float v = vec_mul(srcv[j], vec_splats(id));
+            const __vector int32_t vi = vec_signed(v);
+
+            y[i].qs[4*j + 0] = vec_extract(vi, 0);
+            y[i].qs[4*j + 1] = vec_extract(vi, 1);
+            y[i].qs[4*j + 2] = vec_extract(vi, 2);
+            y[i].qs[4*j + 3] = vec_extract(vi, 3);
+
+            acc = vec_add(acc, vi);
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
+    }
 #else
    GGML_UNUSED(nb);
    // scalar
@@ -2488,6 +2558,37 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
    }

    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#elif defined(__VXE__) || defined(__VXE2__)
+    __vector float acc = vec_splats(0.0f);
+
+    const __vector uint8_t v_m = vec_splats((const uint8_t)0x0F);
+    const __vector int8_t  v_s = vec_splats( (const int8_t)0x08);
+
+    for (; ib < nb; ++ib) {
+        const __vector uint8_t v_x = vec_xl(0, x[ib].qs);
+        const __vector int8_t v_xl = (const __vector int8_t)(v_x & v_m);
+        const __vector int8_t v_xh = (const __vector int8_t)(v_x >> 4);
+
+        const __vector int8_t v_xls = vec_sub(v_xl, v_s);
+        const __vector int8_t v_xhs = vec_sub(v_xh, v_s);
+
+        const __vector int8_t v_yl = vec_xl(0      , y[ib].qs);
+        const __vector int8_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
+
+        const __vector int16_t v_xylso = vec_mulo(v_xls, v_yl);
+        const __vector int16_t v_xylse = vec_mule(v_xls, v_yl);
+        const __vector int16_t v_xyhso = vec_mulo(v_xhs, v_yh);
+        const __vector int16_t v_xyhse = vec_mule(v_xhs, v_yh);
+
+        __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
+
+        const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
+        const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        acc = vec_madd(v_xy, v_d, acc);
+    }
+
+    sumf = acc[0] + acc[1] + acc[2] + acc[3];
 #endif
    for (; ib < nb; ++ib) {
        int sumi0 = 0;
@@ -2781,6 +2882,35 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
    }

    sumf = hsum_float_8(acc) + summs;
+#elif defined(__VXE__) || defined(__VXE2__)
+    float summs = 0;
+    float32x4_t acc = vec_splats(0.0f);
+
+    const uint8x16_t v_m = vec_splat_u8(0x0F);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        const uint8x16_t v_x = vec_xl(0, x[ib].qs);
+        const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
+        const int8x16_t v_xh = (const int8x16_t)(v_x >> 4);
+
+        const int8x16_t v_yl = vec_xl(0      , y[ib].qs);
+        const int8x16_t v_yh = vec_xl(QK8_1/2, y[ib].qs);
+
+        const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
+        const float32x4_t v_xy = vec_float(v_xy_);
+
+        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        acc = vec_madd(v_xy, v_d, acc);
+    }
+
+    sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs;
 #endif
    for (; ib < nb; ++ib) {
        int sumi0 = 0;
@@ -3915,6 +4045,27 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
    }

    sumf = hsum_float_8(acc);
+#elif defined(__VXE__) || defined(__VXE2__)
+    __vector float acc = vec_splats(0.0f);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        const int8x16_t v_xl = vec_xl(0      , x[ib].qs);
+        const int8x16_t v_xh = vec_xl(QK8_0/2, x[ib].qs);
+        const int8x16_t v_yl = vec_xl(0      , y[ib].qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
+
+        const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
+        const float32x4_t v_xy = vec_float(v_xy_);
+        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        acc = vec_madd(v_xy, v_d, acc);
+    }
+
+    sumf = acc[0] + acc[1] + acc[2] + acc[3];
 #endif
    for (; ib < nb; ++ib) {
        int sumi = 0;
@@ -5114,6 +5265,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r

 #if defined(__ARM_FEATURE_SVE)

+    uint32_t aux[3];
    uint32_t utmp[4];

    const int8_t m32 = 32;
@@ -5125,7 +5277,6 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
    const svuint8_t m1_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 1);
    const svuint8_t m2_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 2);
    const svuint8_t m3_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 3);
-    svbool_t pred_s32 = svnot_b_z (svptrue_b32(), svptrue_pat_b32(SV_VL4));

    float sum = 0;

@@ -5138,7 +5289,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
        const int8_t  * restrict q8_sv = y[i].qs;

        // Set up scales
-        uint32_t * aux = &x[i].scales;
+        memcpy(aux, x[i].scales, 12);
        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
@@ -6797,6 +6948,77 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r


    *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
+#elif defined(__VXE__) || defined(__VXE2__)
+    const uint8x16_t v_lm = vec_splat_u8(0x0F);
+    const int32x4_t v_z = vec_splat_s32(0);
+
+    uint8x16_t v_x[2];
+    int8x16_t  v_xl[2];
+    int8x16_t  v_y[2];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
+        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
+        const int16x8_t v_ysums = vec_padd_s16(v_ysumsl, v_ysumsh);
+
+        memcpy(utmp, x[i].scales, 12);
+
+        uint32x4_t v_mins8 = { 0 };
+        v_mins8 = vec_insert(utmp[1] & kmask1, v_mins8, 0);
+        v_mins8 = vec_insert(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), v_mins8, 1);
+
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[0] &= kmask1;
+
+        const int16x8_t v_minsh = (int16x8_t)vec_unpackh((uint8x16_t)v_mins8);
+
+        const int32x4_t v_minso = vec_mulo(v_ysums, v_minsh);
+        const int32x4_t v_minse = vec_mule(v_ysums, v_minsh);
+        const int32x4_t v_mins = v_minso + v_minse;
+        sumf -= dmin * (v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3]);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+        const uint8_t * restrict x0 = x[i].qs;
+        const int8_t  * restrict y0 = y[i].qs;
+
+        int32_t sumi1 = 0;
+        int32_t sumi2 = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            v_x[0] = vec_xl(0 , x0);
+            v_x[1] = vec_xl(16, x0);
+            x0 += 32;
+
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            y0 += 32;
+
+            v_xl[0] = (int8x16_t)vec_and(v_x[0], v_lm);
+            v_xl[1] = (int8x16_t)vec_and(v_x[1], v_lm);
+
+            const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
+            sumi1 += (p1[0] + p1[1] + p1[2] + p1[3]) * scales[2*j+0];
+
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            y0 += 32;
+
+            v_xl[0] = (int8x16_t)vec_sr(v_x[0], 4);
+            v_xl[1] = (int8x16_t)vec_sr(v_x[1], 4);
+
+            const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
+            sumi2 += (p2[0] + p2[1] + p2[2] + p2[3]) * scales[2*j+1];
+        }
+
+        sumf += d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
 #else

    const uint8_t * scales = (const uint8_t*)&utmp[0];
@@ -7526,7 +7748,94 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 4));

    *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
+#elif defined(__VXE__) || defined(__VXE2__)
+    const uint8x16_t v_lm = vec_splat_u8(0x0F);
+    const uint8x16_t v_1m = vec_splat_u8(0x01);
+    const uint8x16_t v_2m = vec_splat_u8(0x02);

+    const int32x4_t v_z = vec_splat_s32(0);
+
+    const uchar8x16_t v_minsm = {
+        0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
+        0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF
+    };
+
+    int8x16_t  q5b[4];
+    uint8x16_t q5h[4];
+
+    uint8x16_t v_xl[2];
+    uint8x16_t v_xh[2];
+    int8x16_t  v_y[4];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
+        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
+        const int16x8_t v_ysums = vec_padd_s16(v_ysumsl, v_ysumsh);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8x16_t v_mins16 = vec_xl(0, (const uint8_t *)utmp);
+        const uint8x16_t v_mins8 = vec_perm(v_mins16, v_mins16, v_minsm);
+        const int16x8_t v_minsh = (int16x8_t)vec_unpackh(v_mins8);
+
+        const int32x4_t v_minsho = vec_mulo(v_ysums, v_minsh);
+        const int32x4_t v_minshe = vec_mule(v_ysums, v_minsh);
+        const int32x4_t v_mins = vec_add(v_minsho, v_minshe);
+        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+        const uint8_t * restrict x0l = x[i].qs;
+        const uint8_t * restrict x0h = x[i].qh;
+        const int8_t  * restrict y0 = y[i].qs;
+
+        v_xh[0] = vec_xl(0 , x0h);
+        v_xh[1] = vec_xl(16, x0h);
+
+        int32_t sumi = 0;
+        for (int j = 0; j < QK_K/64; ++j) {
+            v_xl[0] = vec_xl(0 , x0l);
+            v_xl[1] = vec_xl(16, x0l);
+            x0l += 32;
+
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            v_y[2] = vec_xl(32, y0);
+            v_y[3] = vec_xl(48, y0);
+            y0 += 64;
+
+            q5h[0] = vec_sl(vec_and(v_1m, v_xh[0]), 4);
+            q5h[1] = vec_sl(vec_and(v_1m, v_xh[1]), 4);
+            q5h[2] = vec_sl(vec_and(v_2m, v_xh[0]), 3);
+            q5h[3] = vec_sl(vec_and(v_2m, v_xh[1]), 3);
+            v_xh[0] = vec_sr(v_xh[0], 2);
+            v_xh[1] = vec_sr(v_xh[1], 2);
+
+            q5b[0] = (int8x16_t)vec_or(vec_and(v_xl[0], v_lm), q5h[0]);
+            q5b[1] = (int8x16_t)vec_or(vec_and(v_xl[1], v_lm), q5h[1]);
+            q5b[2] = (int8x16_t)vec_or(vec_sr(v_xl[0], 4), q5h[2]);
+            q5b[3] = (int8x16_t)vec_or(vec_sr(v_xl[1], 4), q5h[3]);
+
+            int32x4_t sumi0 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[0], v_y[0]), q5b[1], v_y[1]);
+            int32x4_t sumi1 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[2], v_y[2]), q5b[3], v_y[3]);
+
+            sumi += (sumi0[0] + sumi0[1] + sumi0[2] + sumi0[3]) * *scales++;
+            sumi += (sumi1[0] + sumi1[1] + sumi1[2] + sumi1[3]) * *scales++;
+        }
+
+        sumf += d * sumi - dmin * mins;
+    }
+
+    *s = sumf;
 #else

    const uint8_t * scales = (const uint8_t*)&utmp[0];
@@ -8243,7 +8552,130 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
    }

    *s = hsum_float_8(acc);
+#elif defined(__VXE__) || defined(__VXE2__)
+    float sum = 0;

+    // Lower 4-bit and upper 2-bit masks
+    const uint8x16_t v_lm = vec_splat_u8(0x0F);
+    const uint8x16_t v_um = vec_splat_u8(0x03);
+
+    const int32x4_t v_z = vec_splat_s32(0);
+
+    int8x16_t  q6b[4];
+    uint8x16_t q6h[4];
+
+    uint8x16_t v_xl[4];
+    uint8x16_t v_xh[2];
+    int8x16_t  v_y[4];
+
+    for (int i = 0; i < nb; ++i) {
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict x0l = x[i].ql;
+        const uint8_t * restrict x0h = x[i].qh;
+        const int8_t  * restrict y0 = y[i].qs;
+
+        const int8_t  * restrict scale = x[i].scales;
+
+        const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
+        const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
+
+        const int8x16_t v_scale  = vec_xl(0, scale);
+        const int16x8_t v_scalel = vec_unpackh(v_scale);
+        const int16x8_t v_scaleh = vec_unpackl(v_scale);
+
+        const int32x4_t v_minslo = vec_mulo(v_ysumsl, v_scalel);
+        const int32x4_t v_minsle = vec_mule(v_ysumsl, v_scalel);
+        const int32x4_t v_minsho = vec_mulo(v_ysumsh, v_scaleh);
+        const int32x4_t v_minshe = vec_mule(v_ysumsh, v_scaleh);
+        const int32x4_t v_mins = v_minslo + v_minsle + v_minsho + v_minshe;
+
+        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+
+        int32_t isum = 0;
+        for (int j = 0; j < QK_K/128; ++j) {
+            // Load model upper 2 bits
+            v_xh[0] = vec_xl(0 , x0h);
+            v_xh[1] = vec_xl(16, x0h);
+            x0h += 32;
+
+            // Load model lower 4 bits
+            v_xl[0] = vec_xl(0 , x0l);
+            v_xl[1] = vec_xl(16, x0l);
+            v_xl[2] = vec_xl(32, x0l);
+            v_xl[3] = vec_xl(48, x0l);
+            x0l += 64;
+
+            // Load activation quants
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            v_y[2] = vec_xl(32, y0);
+            v_y[3] = vec_xl(48, y0);
+            y0 += 64;
+
+            q6h[0] = vec_sl(vec_and(v_um, v_xh[0]), 4);
+            q6h[1] = vec_sl(vec_and(v_um, v_xh[1]), 4);
+            uint8x16_t shifted = vec_sr(v_xh[0], 2);
+            q6h[2] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[1], 2);
+            q6h[3] = vec_sl(vec_and(v_um, shifted), 4);
+
+            q6b[0] = (int8x16_t)(vec_or(vec_and(v_xl[0], v_lm), q6h[0]));
+            q6b[1] = (int8x16_t)(vec_or(vec_and(v_xl[1], v_lm), q6h[1]));
+            q6b[2] = (int8x16_t)(vec_or(vec_and(v_xl[2], v_lm), q6h[2]));
+            q6b[3] = (int8x16_t)(vec_or(vec_and(v_xl[3], v_lm), q6h[3]));
+
+            int32x4_t summs0 = ggml_vec_dot(v_z, q6b[0], v_y[0]);
+            int32x4_t summs1 = ggml_vec_dot(v_z, q6b[1], v_y[1]);
+            int32x4_t summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
+            int32x4_t summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
+
+            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
+                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
+                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
+                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+
+            scale += 4;
+
+
+            // Load activation quants
+            v_y[0] = vec_xl(0 , y0);
+            v_y[1] = vec_xl(16, y0);
+            v_y[2] = vec_xl(32, y0);
+            v_y[3] = vec_xl(48, y0);
+            y0 += 64;
+
+            shifted = vec_sr(v_xh[0], 4);
+            q6h[0] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[1], 4);
+            q6h[1] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[0], 6);
+            q6h[2] = vec_sl(vec_and(v_um, shifted), 4);
+            shifted = vec_sr(v_xh[1], 6);
+            q6h[3] = vec_sl(vec_and(v_um, shifted), 4);
+
+            q6b[0] = (int8x16_t)(vec_or(vec_sr(v_xl[0], 4), q6h[0]));
+            q6b[1] = (int8x16_t)(vec_or(vec_sr(v_xl[1], 4), q6h[1]));
+            q6b[2] = (int8x16_t)(vec_or(vec_sr(v_xl[2], 4), q6h[2]));
+            q6b[3] = (int8x16_t)(vec_or(vec_sr(v_xl[3], 4), q6h[3]));
+
+            summs0 = ggml_vec_dot(v_z, q6b[0], v_y[0]);
+            summs1 = ggml_vec_dot(v_z, q6b[1], v_y[1]);
+            summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
+            summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
+
+            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
+                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
+                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
+                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+
+            scale += 4;
+        }
+
+        sum += d_all * y[i].d * (isum - 32 * mins);
+    }
+
+    *s = sum;
 #else

    int8_t  aux8[QK_K];
@@ -8604,7 +9036,57 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void
    }

    *s = 0.125f * hsum_float_8(accumf);
-
+//#elif defined(__VXE__) || defined(__VXE2__)
+//    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+//
+//    uint32_t aux32[4];
+//    const uint8_t * aux8 = (const uint8_t *)aux32;
+//
+//    float sumf = 0;
+//
+//    for (int i = 0; i < nb; ++i) {
+//        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//        const uint16_t * restrict q2 = x[i].qs;
+//        const int8_t   * restrict q8 = y[i].qs;
+//
+//        float sumf1 = 0, sumf2 = 0;
+//
+//        for (int ib32 = 0; ib32 < QK_K/32; ib += 2) {
+//            int8x16_t q8b0 = vec_xl( 0, q8);
+//            int8x16_t qb81 = vec_xl(16, q8);
+//            int8x16_t q8b2 = vec_xl(32, q8);
+//            int8x16_t q8b3 = vec_xl(48, q8);
+//            q8 += 64;
+//
+//            memcpy(aux32, q2, 4 * sizeof(uint32_t));
+//            q2 += 8;
+//
+//            int8x16_t q2u0 = { *(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1]) };
+//            int8x16_t q2u1 = { *(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3]) };
+//            int8x16_t q2u2 = { *(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9]) };
+//            int8x16_t q2u3 = { *(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11]) };
+//
+//            int8x16_t q2s0 = { *(const int64_t *)(signs64 + ((aux32[1] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >>  7) & 127)) };
+//            int8x16_t q2s1 = { *(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127)) };
+//            int8x16_t q2s2 = { *(const int64_t *)(signs64 + ((aux32[3] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >>  7) & 127)) };
+//            int8x16_t q2s3 = { *(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127)) };
+//
+//            q2u0 = vec_mul(q2u0, q2s0);
+//            q2u1 = vec_mul(q2u1, q2s1);
+//            q2u2 = vec_mul(q2u2, q2s2);
+//            q2u3 = vec_mul(q2u3, q2s3);
+//
+//            const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(vec_splat_s32(0), q2u0, q8b0), q2u1, q8b1);
+//            const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(vec_splat_s32(0), q2u2, q8b2), q2u3, q8b3);
+//
+//            sumf1 += (p1[0] + p1[1] + p1[2] + p1[3]) * (0.5f + (aux32[1] >> 28));
+//            sumf2 += (p2[0] + p2[1] + p2[2] + p2[3]) * (0.5f + (aux32[3] >> 28));
+//        }
+//
+//        sumf += d * (sumf1 + sumf2);
+//    }
+//
+//    *s = 0.25f * sumf;
 #else

    uint32_t aux32[2];
@@ -11365,6 +11847,27 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *

    sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));

+#elif defined(__VXE__) || defined(__VXE2__)
+    const int8x16_t v_k = vec_xl(0, kvalues_iq4nl);
+    const uint8x16_t v_m = vec_splat_u8(0x0F);
+
+    for (; ib < nb; ++ib) {
+        const block_iq4_nl * restrict x0 = &x[ib];
+        const block_q8_0   * restrict y0 = &y[ib];
+
+        const uint8x16_t v_x = vec_xl(0, x0->qs);
+        int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        v_xl = vec_perm(v_k, v_k, (uchar8x16_t)v_xl);
+        v_xh = vec_perm(v_k, v_k, (uchar8x16_t)v_xh);
+
+        const int8x16_t v_yl = vec_xl(0      , y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
+
+        sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
+    }
 #endif
    for (; ib < nb; ++ib) {
        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
@@ -11643,6 +12146,56 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
    }

    *s = hsum_float_8(accum);
+#elif defined(__VXE__) || defined(__VXE2__)
+    const int8x16_t v_k = vec_xl(0, kvalues_iq4nl);
+    const uint8x16_t v_m = vec_splat_u8(0x0F);
+
+    float sumf = 0;
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * restrict q4 = x[ibl].qs;
+        const int8_t  * restrict q8 = y[ibl].qs;
+
+        uint16_t h = x[ibl].scales_h;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/64; ++ib) {
+            const uint8x16_t v_x0 = vec_xl(0       , q4);
+            const uint8x16_t v_x1 = vec_xl(QK4_NL/2, q4);
+            q4 += 32;
+
+            int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+            int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+            int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+            int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+            v_x0l = vec_perm(v_k, v_k, (uchar8x16_t)v_x0l);
+            v_x0h = vec_perm(v_k, v_k, (uchar8x16_t)v_x0h);
+            v_x1l = vec_perm(v_k, v_k, (uchar8x16_t)v_x1l);
+            v_x1h = vec_perm(v_k, v_k, (uchar8x16_t)v_x1h);
+
+            const int8x16_t v_y0 = vec_xl( 0, q8);
+            const int8x16_t v_y1 = vec_xl(16, q8);
+            const int8x16_t v_y2 = vec_xl(32, q8);
+            const int8x16_t v_y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            int32x4_t vsumi0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0l, v_y0), v_x0h, v_y1);
+            int32x4_t vsumi1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1l, v_y2), v_x1h, v_y3);
+
+            int ls1 = ((x[ibl].scales_l[ib] & 0xF) | ((h << 4) & 0x30)) - 32;
+            int ls2 = ((x[ibl].scales_l[ib] >>  4) | ((h << 2) & 0x30)) - 32;
+
+            h >>= 4;
+
+            sumi1 += (vsumi0[0] + vsumi0[1] + vsumi0[2] + vsumi0[3]) * ls1;
+            sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
+        }
+
+        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;

 #else
    float sumf = 0;
@@ -237,6 +237,8 @@ typedef pthread_t ggml_thread_t;
 #else
 #if defined(__POWER9_VECTOR__)
 #define CACHE_LINE_SIZE 128
+#elif defined(__VXE__) || defined(__VXE2__)
+#define CACHE_LINE_SIZE 256
 #else
 #define CACHE_LINE_SIZE 64
 #endif
@@ -1211,6 +1213,87 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F16_VEC_MUL             GGML_F32Cx4_MUL
 #define GGML_F16_VEC_REDUCE          GGML_F32Cx4_REDUCE

+#elif defined(__VXE__) || defined(__VXE2__)
+
+#define GGML_SIMD
+
+// F32 s390x
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR  4
+
+#define GGML_F32x4              __vector float
+#define GGML_F32x4_ZERO         vec_splats(0.0f)
+#define GGML_F32x4_SET1         vec_splats
+#define GGML_F32x4_LOAD(p)      vec_xl(0, p)
+#define GGML_F32x4_STORE(p, r)  vec_xst(r, 0, p)
+#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
+#define GGML_F32x4_ADD          vec_add
+#define GGML_F32x4_MUL          vec_mul
+#define GGML_F32x4_REDUCE(res, x)                   \
+{                                                   \
+    int offset = GGML_F32_ARR >> 1;                 \
+    for (int i = 0; i < offset; ++i) {              \
+        x[i] = vec_add(x[i], x[offset + i]);        \
+    }                                               \
+    offset >>= 1;                                   \
+    for (int i = 0; i < offset; ++i) {              \
+        x[i] = vec_add(x[i], x[offset + i]);        \
+    }                                               \
+    offset >>= 1;                                   \
+    for (int i = 0; i < offset; ++i) {              \
+        x[i] = vec_add(x[i], x[offset + i]);        \
+    }                                               \
+    res = vec_extract(x[0], 0) +                    \
+          vec_extract(x[0], 1) +                    \
+          vec_extract(x[0], 2) +                    \
+          vec_extract(x[0], 3);                     \
+}
+
+#define GGML_F32_VEC        GGML_F32x4
+#define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 s390x
+#define GGML_F16_STEP GGML_F32_STEP
+#define GGML_F16_EPR  GGML_F32_EPR
+
+static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
+    float tmp[4];
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+
+    return vec_xl(0, tmp);
+}
+
+static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
+    float arr[4];
+
+    vec_xst(y, 0, arr);
+
+    for (int i = 0; i < 4; i++) {
+        x[i] = GGML_FP32_TO_FP16(arr[i]);
+    }
+}
+
+#define GGML_F16_VEC                GGML_F32x4
+#define GGML_F16_VEC_ZERO           GGML_F32x4_ZERO
+#define GGML_F16_VEC_SET1           GGML_F32x4_SET1
+#define GGML_F16_VEC_LOAD(p, i)     __lzs_f16cx4_load(p)
+#define GGML_F16_VEC_STORE(p, r, i) __lzs_f16cx4_store(p, r[i])
+#define GGML_F16_VEC_FMA            GGML_F32x4_FMA
+#define GGML_F16_VEC_ADD            GGML_F32x4_ADD
+#define GGML_F16_VEC_MUL            GGML_F32x4_MUL
+#define GGML_F16_VEC_REDUCE         GGML_F32x4_REDUCE
+
 #endif

 // GGML_F32_ARR / GGML_F16_ARR
@@ -14419,6 +14502,14 @@ int ggml_cpu_has_vsx(void) {
 #endif
 }

+int ggml_cpu_has_vxe(void) {
+#if defined(__VXE__) || defined(__VXE2__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
    return ggml_arm_arch_features.has_neon;
@@ -557,6 +557,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
        if (ggml_cpu_has_vsx()) {
            features.push_back({ "VSX", "1" });
        }
+        if (ggml_cpu_has_vxe()) {
+            features.push_back({ "VXE", "1" });
+        }
        if (ggml_cpu_has_wasm_simd()) {
            features.push_back({ "WASM_SIMD", "1" });
        }
@@ -69,6 +69,10 @@ if (CUDAToolkit_FOUND)
        add_compile_definitions(GGML_CUDA_NO_VMM)
    endif()

+    if (NOT GGML_CUDA_FA)
+        add_compile_definitions(GGML_CUDA_NO_FA)
+    endif()
+
    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
        add_compile_definitions(GGML_CUDA_F16)
    endif()
@@ -204,9 +204,9 @@ typedef float2 dfloat2;
 #define CP_ASYNC_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE

-#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
 #define FLASH_ATTN_AVAILABLE
-#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
+#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)

 static bool fp16_available(const int cc) {
    return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
@@ -24,7 +24,7 @@ static __device__ __forceinline__ void cp_async_cg_16(const unsigned int dst, co
    } else
 #endif // CUDART_VERSION >= 11040
    {
-        asm volatile("cp.async.cg.shared.global.L2 [%0], [%1], 16;"
+        asm volatile("cp.async.cg.shared.global [%0], [%1], 16;"
            : : "r"(dst), "l"(src));
    }
 #else
@@ -516,27 +516,25 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
        nullptr;
 }

-// The HIP compiler for some reason complains that it can't unroll a loop because of the jt*ncols + j >= ne01 conditional.
-#ifdef __clang__
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wpass-failed"
-#endif // __clang__
-
-template<int D, int ncols, int KQ_stride> // D == head size
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
+template<int D, int ncols1, int ncols2, int KQ_stride> // D == head size
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_stream_k_fixup(
        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne11) {
-    const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols);
-
-    const int iter_k = ne11 / KQ_stride;
-    const int iter_j = (ne01 + (ncols - 1)) / ncols;
+    constexpr int ncols = ncols1*ncols2;

    const int bidx0 = blockIdx.x;
+    const int j     = blockIdx.y;
+    const int c     = blockIdx.z;
+    const int jc    = j*ncols2 + c;
+    const int tid   = threadIdx.x;

-    const int kbc0      = (bidx0 + 0)*iter_k*iter_j*ne02 / gridDim.x;
-    const int kbc0_stop = (bidx0 + 1)*iter_k*iter_j*ne02 / gridDim.x;
+    const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols);
+
+    const int iter_k = ne11 / FATTN_KQ_STRIDE;
+    const int iter_j = (ne01 + (ncols1 - 1)) / ncols1;
+
+    const int kbc0      = (bidx0 + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+    const int kbc0_stop = (bidx0 + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;

    const bool did_not_have_any_data   = kbc0 == kbc0_stop;
    const bool wrote_beginning_of_tile = kbc0 % iter_k == 0;
@@ -548,22 +546,22 @@ static __global__ void flash_attn_stream_k_fixup(
    const int channel = kbc0 / (iter_k*iter_j);
    const int jt      = (kbc0 - channel*iter_k*iter_j) / iter_k;

-    dst += jt*ncols*ne02*D + channel*D;
+    if (jt*ncols1 + j >= ne01) {
+        return;
+    }
+
+    dst += jt*ne02*(ncols1*D) + channel*(ncols2*D) + (j*ne02 + c)*D + tid;

    // Load the partial result that needs a fixup:
-    float dst_val[ncols] = {0.0f};
-    float max_val[ncols] = {0.0f};
-    float rowsum[ncols]  = {0.0f};
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (jt*ncols + j >= ne01) {
-            break;
-        }
-        dst_val[j] = dst[j*ne02*D + threadIdx.x];
+    float dst_val = 0.0f;
+    float max_val = 0.0f;
+    float rowsum  = 0.0f;
+    {
+        dst_val = *dst;

-        const float2 tmp = dst_fixup[bidx0*ncols + j];
-        max_val[j] = tmp.x;
-        rowsum[j]  = tmp.y;
+        const float2 tmp = dst_fixup[bidx0*ncols + jc];
+        max_val = tmp.x;
+        rowsum  = tmp.y;
    }

    // Iterate over previous blocks and compute the combined results.
@@ -571,36 +569,30 @@ static __global__ void flash_attn_stream_k_fixup(
    int bidx = bidx0 - 1;
    int kbc_stop = kbc0;
    while(true) {
-        const int kbc = bidx*iter_k*iter_j*ne02 / gridDim.x;
+        const int kbc = bidx*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
        if (kbc == kbc_stop) { // Did not have any data.
            bidx--;
            kbc_stop = kbc;
            continue;
        }

-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            if (jt*ncols + j >= ne01) {
-                break;
-            }
-            const float dst_add = dst_fixup_data[bidx*ncols*D + j*D + threadIdx.x];
+        const float dst_add = dst_fixup_data[bidx*ncols*D + jc*D + tid];

-            const float2 tmp = dst_fixup[(gridDim.x + bidx)*ncols + j];
+        const float2 tmp = dst_fixup[(gridDim.x + bidx)*ncols + jc];

-            // Scale the current and new value accumulators depending on the max. values.
-            const float max_val_new = fmaxf(max_val[j], tmp.x);
+        // Scale the current and new value accumulators depending on the max. values.
+        const float max_val_new = fmaxf(max_val, tmp.x);

-            const float diff_val = max_val[j] - max_val_new;
-            const float diff_add = tmp.x      - max_val_new;
+        const float diff_val = max_val - max_val_new;
+        const float diff_add = tmp.x   - max_val_new;

-            const float scale_val = diff_val >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_val) : 0.0f;
-            const float scale_add = diff_add >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_add) : 0.0f;
+        const float scale_val = diff_val >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_val) : 0.0f;
+        const float scale_add = diff_add >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_add) : 0.0f;

-            dst_val[j] = scale_val*dst_val[j] + scale_add*dst_add;
-            rowsum[j]  = scale_val*rowsum[j]  + scale_add*tmp.y;
+        dst_val = scale_val*dst_val + scale_add*dst_add;
+        rowsum  = scale_val*rowsum  + scale_add*tmp.y;

-            max_val[j] = max_val_new;
-        }
+        max_val = max_val_new;

        // If this block started in a previous tile we are done and don't need to combine additional partial results.
        if (kbc % iter_k == 0 || kbc/iter_k < kbc0/iter_k) {
@@ -611,19 +603,9 @@ static __global__ void flash_attn_stream_k_fixup(
    }

    // Write back final result:
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (jt*ncols + j >= ne01) {
-            return;
-        }
-        dst[j*ne02*D + threadIdx.x] = dst_val[j] / rowsum[j];
-    }
+    *dst = dst_val / rowsum;
 }

-#ifdef __clang__
-#pragma clang diagnostic pop
-#endif // __clang__
-
 template<int D, int parallel_blocks> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
@@ -690,11 +672,13 @@ static void on_no_fattn_vec_case(const int D) {
 }

 // parallel_blocks == 0 is stream-k decomposition
-template <int D, int cols_per_block, int parallel_blocks, int KQ_stride>
+template <int D, int ncols1, int ncols2, int parallel_blocks, int KQ_stride>
 void launch_fattn(
    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel,
    const int nwarps, const size_t nbytes_shared, const bool need_f16_K, const bool need_f16_V
 ) {
+    constexpr int ncols = ncols1 * ncols2;
+
    const ggml_tensor * Q = dst->src[0];
    const ggml_tensor * K = dst->src[1];
    const ggml_tensor * V = dst->src[2];
@@ -763,25 +747,26 @@ void launch_fattn(
        nb23 = nb23*bs*sizeof(half)/ts;
    }

-    const int ntiles_x = ((Q->ne[1] + cols_per_block - 1) / cols_per_block);
-    const int ntiles_total = ntiles_x*Q->ne[2]*Q->ne[3];
+    const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
+    const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];

    const dim3 block_dim(WARP_SIZE, nwarps, 1);
    dim3 blocks_num;
    if (parallel_blocks == 0) {
        // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
-        const int tiles_nwaves  = (ntiles_total + 2*nsm - 1) / (2*nsm);
-        const int tiles_efficiency_percent = 100 * ntiles_total / (2*nsm*tiles_nwaves);
+        const int max_blocks = 2*nsm;
+        const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks;
+        const int tiles_efficiency_percent = 100 * ntiles_total / (max_blocks*tiles_nwaves);

-        const int nblocks_stream_k = 2*nsm;
+        const int nblocks_stream_k = max_blocks;

-        const bool use_stream_k = tiles_efficiency_percent < 75 || cc >= GGML_CUDA_CC_ADA_LOVELACE;
+        const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || tiles_efficiency_percent < 75;

        blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_total;
        blocks_num.y = 1;
        blocks_num.z = 1;

-        dst_tmp_meta.alloc(blocks_num.x*cols_per_block * (2*2 + D) * sizeof(float));
+        dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + D) * sizeof(float));
    } else {
        blocks_num.x = parallel_blocks*ntiles_x;
        blocks_num.y = Q->ne[2];
@@ -793,7 +778,6 @@ void launch_fattn(
        }
    }

-
    float scale         = 1.0f;
    float max_bias      = 0.0f;
    float logit_softcap = 0.0f;
@@ -832,9 +816,9 @@ void launch_fattn(
    if constexpr (parallel_blocks == 0) {
        if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
            const dim3 block_dim_combine(D, 1, 1);
-            const dim3 blocks_num_combine = blocks_num;
+            const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2};

-            flash_attn_stream_k_fixup<D, cols_per_block, KQ_stride>
+            flash_attn_stream_k_fixup<D, ncols1, ncols2, KQ_stride>
                <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]);
        }
@@ -44,12 +44,7 @@ static __global__ void flash_attn_tile_ext_f16(
        const int ne1,
        const int ne2,
        const int ne3) {
-#ifdef FP16_AVAILABLE
-
-#ifndef FLASH_ATTN_AVAILABLE
-    NO_DEVICE_CODE;
-    return;
-#endif // FLASH_ATTN_AVAILABLE
+#if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)

    // Skip unused kernel variants for faster compilation:
 #ifdef FP16_MMA_AVAILABLE
@@ -290,7 +285,7 @@ static __global__ void flash_attn_tile_ext_f16(
    }
 #else
   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
+#endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }

 template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
@@ -302,14 +297,14 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
        } break;
        case 128: {
            constexpr int    D             = 128;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
        } break;
        default: {
            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -44,10 +44,7 @@ static __global__ void flash_attn_tile_ext_f32(
        const int ne1,
        const int ne2,
        const int ne3) {
-#ifndef FLASH_ATTN_AVAILABLE
-    NO_DEVICE_CODE;
-    return;
-#endif // FLASH_ATTN_AVAILABLE
+#ifdef FLASH_ATTN_AVAILABLE

    // Skip unused kernel variants for faster compilation:
 #ifdef FP16_MMA_AVAILABLE
@@ -285,6 +282,9 @@ static __global__ void flash_attn_tile_ext_f32(
            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
        }
    }
+#else
+    NO_DEVICE_CODE;
+#endif // FLASH_ATTN_AVAILABLE
 }

 template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
@@ -296,14 +296,14 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
        } break;
        case 128: {
            constexpr int    D             = 128;
            constexpr int    nwarps        = 8;
            constexpr size_t nbytes_shared = 0;
            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
-            launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
        } break;
        default: {
            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -41,12 +41,7 @@ static __global__ void flash_attn_vec_ext_f16(
        const int ne1,
        const int ne2,
        const int ne3) {
-#ifdef FP16_AVAILABLE
-
-#ifndef FLASH_ATTN_AVAILABLE
-    NO_DEVICE_CODE;
-    return;
-#endif // FLASH_ATTN_AVAILABLE
+#if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)

    // Skip unused kernel variants for faster compilation:
    if (use_logit_softcap && !(D == 128 || D == 256)) {
@@ -300,7 +295,7 @@ static __global__ void flash_attn_vec_ext_f16(
    }
 #else
   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
+#endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }

 template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
@@ -310,7 +305,7 @@ void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx,
    constexpr bool need_f16_K = D != 128;
    constexpr bool need_f16_V = D != 128 && D != 64;
    constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
 }

 template <int D, ggml_type type_K, ggml_type type_V>
@@ -41,10 +41,7 @@ static __global__ void flash_attn_vec_ext_f32(
        const int ne1,
        const int ne2,
        const int ne3) {
-#ifndef FLASH_ATTN_AVAILABLE
-    NO_DEVICE_CODE;
-    return;
-#endif // FLASH_ATTN_AVAILABLE
+#ifdef FLASH_ATTN_AVAILABLE

    // Skip unused kernel variants for faster compilation:
    if (use_logit_softcap && !(D == 128 || D == 256)) {
@@ -281,6 +278,9 @@ static __global__ void flash_attn_vec_ext_f32(
    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
    }
+#else
+    NO_DEVICE_CODE;
+#endif // FLASH_ATTN_AVAILABLE
 }

 template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
@@ -290,7 +290,7 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx,
    constexpr bool need_f16_K = D != 128;
    constexpr bool need_f16_V = D != 128 && D != 64;
    constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
 }

 template <int D, ggml_type type_K, ggml_type type_V>
@@ -51,7 +51,7 @@ static __global__ void flash_attn_ext_f16(
        const int ne1,
        const int ne2,
        const int ne3) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(FLASH_ATTN_AVAILABLE) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
    // Skip unused kernel variants for faster compilation:
    if (use_logit_softcap && !(D == 128 || D == 256)) {
        NO_DEVICE_CODE;
@@ -425,7 +425,7 @@ static __global__ void flash_attn_ext_f16(
    }
 #else
   NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(FLASH_ATTN_AVAILABLE) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 }

 constexpr int get_max_power_of_2(int x) {
@@ -478,7 +478,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm
            fattn_kernel = flash_attn_ext_f16<
                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
        }
-        launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true);
+        launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true);
        return;
    }
    if (2*blocks_num_pb1 < 2*nsm) {
@@ -493,7 +493,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm
            fattn_kernel = flash_attn_ext_f16<
                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
        }
-        launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true);
+        launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true);
        return;
    }
    constexpr int parallel_blocks = 1;
@@ -507,7 +507,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm
        fattn_kernel = flash_attn_ext_f16<
            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
    }
-    launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true);
+    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true);
 }

 void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -8,28 +8,50 @@
 #include "fattn-wmma-f16.cuh"
 #include "fattn.cuh"

-template <int cols_per_block>
+template <int D, int ncols2>
+static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+
+    if (Q->ne[1] <= 8/ncols2) {
+        ggml_cuda_flash_attn_ext_mma_f16_case<D, 8/ncols2, ncols2>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 16/ncols2) {
+        ggml_cuda_flash_attn_ext_mma_f16_case<D, 16/ncols2, ncols2>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32/ncols2) {
+        ggml_cuda_flash_attn_ext_mma_f16_case<D, 32/ncols2, ncols2>(ctx, dst);
+        return;
+    }
+
+    ggml_cuda_flash_attn_ext_mma_f16_case<D, 64/ncols2, ncols2>(ctx, dst);
+}
+
+template <int ncols2>
 static void ggml_cuda_flash_attn_ext_mma_f16_switch_hs(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * Q = dst->src[0];

    switch (Q->ne[0]) {
        case 64:
-            ggml_cuda_flash_attn_ext_mma_f16_case< 64, cols_per_block>(ctx, dst);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 64, ncols2>(ctx, dst);
            break;
        case 80:
-            ggml_cuda_flash_attn_ext_mma_f16_case< 80, cols_per_block>(ctx, dst);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 80, ncols2>(ctx, dst);
            break;
        case 96:
-            ggml_cuda_flash_attn_ext_mma_f16_case< 96, cols_per_block>(ctx, dst);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 96, ncols2>(ctx, dst);
            break;
        case 112:
-            ggml_cuda_flash_attn_ext_mma_f16_case<112, cols_per_block>(ctx, dst);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<112, ncols2>(ctx, dst);
            break;
        case 128:
-            ggml_cuda_flash_attn_ext_mma_f16_case<128, cols_per_block>(ctx, dst);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<128, ncols2>(ctx, dst);
            break;
        case 256:
-            ggml_cuda_flash_attn_ext_mma_f16_case<256, cols_per_block>(ctx, dst);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<256, ncols2>(ctx, dst);
            break;
        default:
            GGML_ABORT("fatal error");
@@ -38,24 +60,35 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_hs(ggml_backend_cuda_context
 }

 static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * KQV  = dst;
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * mask = dst->src[3];

-    if (Q->ne[1] <= 8) {
+    float max_bias = 0.0f;
+    memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+    const float use_gqa_opt = mask && max_bias == 0.0f;
+
+    GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+    const int gqa_ratio = Q->ne[2] / K->ne[2];
+
+    if (use_gqa_opt && gqa_ratio % 8 == 0) {
        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<8>(ctx, dst);
        return;
    }

-    if (Q->ne[1] <= 16) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<16>(ctx, dst);
+    if (use_gqa_opt && gqa_ratio == 4) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<4>(ctx, dst);
        return;
    }

-    if (Q->ne[1] <= 32) {
-        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<32>(ctx, dst);
+    if (use_gqa_opt && gqa_ratio == 2) {
+        ggml_cuda_flash_attn_ext_mma_f16_switch_hs<2>(ctx, dst);
        return;
    }

-    ggml_cuda_flash_attn_ext_mma_f16_switch_hs<64>(ctx, dst);
+    ggml_cuda_flash_attn_ext_mma_f16_switch_hs<1>(ctx, dst);
 }

 #define FATTN_VEC_F16_CASE(D, type_K, type_V)                               \
@@ -209,8 +242,11 @@ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, gg
 }

 void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * KQV  = dst;
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];

    ggml_cuda_set_device(ctx.device);
    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
@@ -252,7 +288,10 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
        return;
    }

-    if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
+    const int gqa_ratio = Q->ne[2] / K->ne[2];
+    const bool mma_fast_for_bs1 = fp16_mma_available(cc) && gqa_ratio % 2 == 0 &&
+        K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16 && mask;
+    if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0 && !mma_fast_for_bs1) {
        if (prec == GGML_PREC_DEFAULT) {
            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
            return;
@@ -3203,7 +3203,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
        case GGML_OP_FLASH_ATTN_EXT: {
 #ifndef FLASH_ATTN_AVAILABLE
            return false;
-#endif
+#endif // FLASH_ATTN_AVAILABLE
            if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                return false;
            }
@@ -73,6 +73,8 @@ namespace ggml_cuda_mma {
                return threadIdx.x / 4;
            } else if constexpr (I == 16 && J == 8) {
                return (l / 2) * 8 + threadIdx.x / 4;
+            } else if constexpr (I == 16 && J == 16) {
+                return ((l / 2) % 2) * 8 + threadIdx.x / 4;
            } else {
                static_assert(I == -1 && J == -1, "template specialization not implemented");
            }
@@ -85,6 +87,8 @@ namespace ggml_cuda_mma {
                return 4 * l + threadIdx.x % 4;
            } else if constexpr (I == 16 && J == 8) {
                return 2 * (threadIdx.x % 4) + l % 2;
+            } else if constexpr (I == 16 && J == 16) {
+                return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2;
            } else {
                static_assert(I == -1 && J == -1, "template specialization not implemented");
            }
@@ -289,6 +293,42 @@ namespace ggml_cuda_mma {
 #endif // NEW_MMA_AVAILABLE
    }

+    static __device__ __forceinline__ void mma(
+            tile<16, 8, half2> & D, const tile<16, 8, half2> & A, const tile<16, 8, half2> & B) {
+#ifdef NEW_MMA_AVAILABLE
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+        asm("mma.sync.aligned.m16n8k16.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3, %4, %5}, {%6, %7}, {%0, %1};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[2]));
+        asm("mma.sync.aligned.m16n8k16.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3, %4, %5}, {%6, %7}, {%0, %1};"
+            : "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]), "r"(Bxi[3]));
+#else
+        // On Turing m16n8k16 mma is not available, use 4x m8n8k8 mma instead:
+        asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]));
+        asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]));
+        asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};"
+            : "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};"
+            : "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif // NEW_MMA_AVAILABLE
+    }
+
    static __device__ __forceinline__ void mma(
            tile<16, 8, float> & D, const tile<16, 8, half2> & A, const tile<8, 8, half2> & B) {
 #ifdef NEW_MMA_AVAILABLE
@@ -316,4 +356,39 @@ namespace ggml_cuda_mma {
 #endif // NEW_MMA_AVAILABLE
    }

+    static __device__ __forceinline__ void mma(
+            tile<16, 16, float> & D, const tile<16, 8, half2> & A, const tile<16, 8, half2> & B) {
+#ifdef NEW_MMA_AVAILABLE
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+        asm("mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[2]));
+        asm("mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]), "r"(Bxi[3]));
+#else
+        // On Turing m16n8k16 mma is not available, use 4x m8n8k8 mma instead:
+        asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]));
+        asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]));
+        asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#else
+        GGML_UNUSED(D);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        NO_DEVICE_CODE;
+#endif // NEW_MMA_AVAILABLE
+    }
 }
@@ -1,10 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-mma-f16.cuh"
-
-DECL_FATTN_MMA_F16_CASE(64, 16);
-DECL_FATTN_MMA_F16_CASE(80, 16);
-DECL_FATTN_MMA_F16_CASE(96, 16);
-DECL_FATTN_MMA_F16_CASE(112, 16);
-DECL_FATTN_MMA_F16_CASE(128, 16);
-DECL_FATTN_MMA_F16_CASE(256, 16);
@@ -1,10 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-mma-f16.cuh"
-
-DECL_FATTN_MMA_F16_CASE(64, 32);
-DECL_FATTN_MMA_F16_CASE(80, 32);
-DECL_FATTN_MMA_F16_CASE(96, 32);
-DECL_FATTN_MMA_F16_CASE(112, 32);
-DECL_FATTN_MMA_F16_CASE(128, 32);
-DECL_FATTN_MMA_F16_CASE(256, 32);
@@ -1,10 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-mma-f16.cuh"
-
-DECL_FATTN_MMA_F16_CASE(64, 64);
-DECL_FATTN_MMA_F16_CASE(80, 64);
-DECL_FATTN_MMA_F16_CASE(96, 64);
-DECL_FATTN_MMA_F16_CASE(112, 64);
-DECL_FATTN_MMA_F16_CASE(128, 64);
-DECL_FATTN_MMA_F16_CASE(256, 64);
@@ -1,10 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-mma-f16.cuh"
-
-DECL_FATTN_MMA_F16_CASE(64, 8);
-DECL_FATTN_MMA_F16_CASE(80, 8);
-DECL_FATTN_MMA_F16_CASE(96, 8);
-DECL_FATTN_MMA_F16_CASE(112, 8);
-DECL_FATTN_MMA_F16_CASE(128, 8);
-DECL_FATTN_MMA_F16_CASE(256, 8);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 1, 8);
+DECL_FATTN_MMA_F16_CASE(80, 1, 8);
+DECL_FATTN_MMA_F16_CASE(96, 1, 8);
+DECL_FATTN_MMA_F16_CASE(112, 1, 8);
+DECL_FATTN_MMA_F16_CASE(128, 1, 8);
+DECL_FATTN_MMA_F16_CASE(256, 1, 8);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 16, 1);
+DECL_FATTN_MMA_F16_CASE(80, 16, 1);
+DECL_FATTN_MMA_F16_CASE(96, 16, 1);
+DECL_FATTN_MMA_F16_CASE(112, 16, 1);
+DECL_FATTN_MMA_F16_CASE(128, 16, 1);
+DECL_FATTN_MMA_F16_CASE(256, 16, 1);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 16, 2);
+DECL_FATTN_MMA_F16_CASE(80, 16, 2);
+DECL_FATTN_MMA_F16_CASE(96, 16, 2);
+DECL_FATTN_MMA_F16_CASE(112, 16, 2);
+DECL_FATTN_MMA_F16_CASE(128, 16, 2);
+DECL_FATTN_MMA_F16_CASE(256, 16, 2);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 16, 4);
+DECL_FATTN_MMA_F16_CASE(80, 16, 4);
+DECL_FATTN_MMA_F16_CASE(96, 16, 4);
+DECL_FATTN_MMA_F16_CASE(112, 16, 4);
+DECL_FATTN_MMA_F16_CASE(128, 16, 4);
+DECL_FATTN_MMA_F16_CASE(256, 16, 4);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 2, 4);
+DECL_FATTN_MMA_F16_CASE(80, 2, 4);
+DECL_FATTN_MMA_F16_CASE(96, 2, 4);
+DECL_FATTN_MMA_F16_CASE(112, 2, 4);
+DECL_FATTN_MMA_F16_CASE(128, 2, 4);
+DECL_FATTN_MMA_F16_CASE(256, 2, 4);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 2, 8);
+DECL_FATTN_MMA_F16_CASE(80, 2, 8);
+DECL_FATTN_MMA_F16_CASE(96, 2, 8);
+DECL_FATTN_MMA_F16_CASE(112, 2, 8);
+DECL_FATTN_MMA_F16_CASE(128, 2, 8);
+DECL_FATTN_MMA_F16_CASE(256, 2, 8);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 32, 1);
+DECL_FATTN_MMA_F16_CASE(80, 32, 1);
+DECL_FATTN_MMA_F16_CASE(96, 32, 1);
+DECL_FATTN_MMA_F16_CASE(112, 32, 1);
+DECL_FATTN_MMA_F16_CASE(128, 32, 1);
+DECL_FATTN_MMA_F16_CASE(256, 32, 1);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 32, 2);
+DECL_FATTN_MMA_F16_CASE(80, 32, 2);
+DECL_FATTN_MMA_F16_CASE(96, 32, 2);
+DECL_FATTN_MMA_F16_CASE(112, 32, 2);
+DECL_FATTN_MMA_F16_CASE(128, 32, 2);
+DECL_FATTN_MMA_F16_CASE(256, 32, 2);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 4, 2);
+DECL_FATTN_MMA_F16_CASE(80, 4, 2);
+DECL_FATTN_MMA_F16_CASE(96, 4, 2);
+DECL_FATTN_MMA_F16_CASE(112, 4, 2);
+DECL_FATTN_MMA_F16_CASE(128, 4, 2);
+DECL_FATTN_MMA_F16_CASE(256, 4, 2);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 4, 4);
+DECL_FATTN_MMA_F16_CASE(80, 4, 4);
+DECL_FATTN_MMA_F16_CASE(96, 4, 4);
+DECL_FATTN_MMA_F16_CASE(112, 4, 4);
+DECL_FATTN_MMA_F16_CASE(128, 4, 4);
+DECL_FATTN_MMA_F16_CASE(256, 4, 4);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 4, 8);
+DECL_FATTN_MMA_F16_CASE(80, 4, 8);
+DECL_FATTN_MMA_F16_CASE(96, 4, 8);
+DECL_FATTN_MMA_F16_CASE(112, 4, 8);
+DECL_FATTN_MMA_F16_CASE(128, 4, 8);
+DECL_FATTN_MMA_F16_CASE(256, 4, 8);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 64, 1);
+DECL_FATTN_MMA_F16_CASE(80, 64, 1);
+DECL_FATTN_MMA_F16_CASE(96, 64, 1);
+DECL_FATTN_MMA_F16_CASE(112, 64, 1);
+DECL_FATTN_MMA_F16_CASE(128, 64, 1);
+DECL_FATTN_MMA_F16_CASE(256, 64, 1);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 8, 1);
+DECL_FATTN_MMA_F16_CASE(80, 8, 1);
+DECL_FATTN_MMA_F16_CASE(96, 8, 1);
+DECL_FATTN_MMA_F16_CASE(112, 8, 1);
+DECL_FATTN_MMA_F16_CASE(128, 8, 1);
+DECL_FATTN_MMA_F16_CASE(256, 8, 1);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 8, 2);
+DECL_FATTN_MMA_F16_CASE(80, 8, 2);
+DECL_FATTN_MMA_F16_CASE(96, 8, 2);
+DECL_FATTN_MMA_F16_CASE(112, 8, 2);
+DECL_FATTN_MMA_F16_CASE(128, 8, 2);
+DECL_FATTN_MMA_F16_CASE(256, 8, 2);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 8, 4);
+DECL_FATTN_MMA_F16_CASE(80, 8, 4);
+DECL_FATTN_MMA_F16_CASE(96, 8, 4);
+DECL_FATTN_MMA_F16_CASE(112, 8, 4);
+DECL_FATTN_MMA_F16_CASE(128, 8, 4);
+DECL_FATTN_MMA_F16_CASE(256, 8, 4);
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(64, 8, 8);
+DECL_FATTN_MMA_F16_CASE(80, 8, 8);
+DECL_FATTN_MMA_F16_CASE(96, 8, 8);
+DECL_FATTN_MMA_F16_CASE(112, 8, 8);
+DECL_FATTN_MMA_F16_CASE(128, 8, 8);
+DECL_FATTN_MMA_F16_CASE(256, 8, 8);
@@ -18,7 +18,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f

 """

-SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {cols_per_block});\n"
+SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {ncols1}, {ncols2});\n"

 TYPES_MMQ = [
    "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
@@ -57,12 +57,18 @@ for vkq_size in [16, 32]:
                with open(f"fattn-vec-f{vkq_size}-instance-hs{head_size}-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f:
                    f.write(SOURCE_FATTN_VEC.format(vkq_size=vkq_size, head_size=head_size, type_k=type_k, type_v=type_v))

-for cols_per_block in [8, 16, 32, 64]:
-    with open(f"fattn-mma-f16-instance-cpb{cols_per_block}.cu", "w") as f:
-        f.write(SOURCE_FATTN_MMA_START)
+for ncols in [8, 16, 32, 64, 128]:
+    for ncols2 in [1, 2, 4, 8]:
+        ncols1 = ncols // ncols2
+        if ncols == 128:
+            continue  # Too much register pressure.
+        with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f:
+            f.write(SOURCE_FATTN_MMA_START)

-        for head_size in [64, 80, 96, 112, 128, 256]:
-            f.write(SOURCE_FATTN_MMA_CASE.format(cols_per_block=cols_per_block, head_size=head_size))
+            for head_size in [64, 80, 96, 112, 128, 256]:
+                if ncols == 128 and head_size == 256:
+                    continue  # Needs too much shared memory.
+                f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size=head_size))

 for type in TYPES_MMQ:
    with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f:
@@ -107,6 +107,10 @@ if (GGML_HIP_NO_VMM)
    add_compile_definitions(GGML_HIP_NO_VMM)
 endif()

+if (NOT GGML_CUDA_FA)
+    add_compile_definitions(GGML_CUDA_NO_FA)
+endif()
+
 if (CXX_IS_HIPCC)
    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
    target_link_libraries(ggml-hip PRIVATE hip::device)
@@ -407,6 +407,16 @@ enum ggml_metal_kernel_type {
    GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,
    GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,
    GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,
+    GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F32,
+    GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F16,
+    GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F32,
+    GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F16,
+    GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F32,
+    GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F16,
+    GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F32,
+    GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F16,
+    GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F16,
    GGML_METAL_KERNEL_TYPE_CONCAT,
    GGML_METAL_KERNEL_TYPE_SQR,
    GGML_METAL_KERNEL_TYPE_SQRT,
@@ -1012,6 +1022,16 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,                  cpy_f32_q5_0,                   true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,                  cpy_f32_q5_1,                   true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,                cpy_f32_iq4_nl,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F32,                  cpy_q4_0_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F16,                  cpy_q4_0_f16,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F32,                  cpy_q4_1_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F16,                  cpy_q4_1_f16,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F32,                  cpy_q5_0_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F16,                  cpy_q5_0_f16,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F32,                  cpy_q5_1_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F16,                  cpy_q5_1_f16,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F32,                  cpy_q8_0_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F16,                  cpy_q8_0_f16,                   true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                        concat,                         true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                           sqr,                            true);
        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT,                          sqrt,                           true);
@@ -1287,6 +1307,18 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                            default:
                                return false;
                        }
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                        switch (op->type) {
+                            case GGML_TYPE_F32:
+                            case GGML_TYPE_F16:
+                                return true;
+                            default:
+                                return false;
+                        }
                    default:
                        return false;
                };
@@ -3899,10 +3931,6 @@ static void ggml_metal_encode_node(
        case GGML_OP_CPY:
        case GGML_OP_CONT:
            {
-                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
-
-                int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
-
                id<MTLComputePipelineState> pipeline = nil;

                switch (src0t) {
@@ -3936,7 +3964,47 @@ static void ggml_metal_encode_node(
                            switch (dstt) {
                                case GGML_TYPE_F32:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_BF16_F32].pipeline; break;
                                case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16].pipeline; break;
-                                default: GGML_ASSERT(false && "not implemented");
+                                default: GGML_ABORT("not implemented");
+                            };
+                        } break;
+                    case GGML_TYPE_Q4_0:
+                        {
+                            switch (dstt) {
+                                case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F32].pipeline; break;
+                                case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F16].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            };
+                        } break;
+                    case GGML_TYPE_Q4_1:
+                        {
+                            switch (dstt) {
+                                case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F32].pipeline; break;
+                                case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F16].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            };
+                        } break;
+                    case GGML_TYPE_Q5_0:
+                        {
+                            switch (dstt) {
+                                case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F32].pipeline; break;
+                                case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F16].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            };
+                        } break;
+                    case GGML_TYPE_Q5_1:
+                        {
+                            switch (dstt) {
+                                case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F32].pipeline; break;
+                                case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F16].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            };
+                        } break;
+                    case GGML_TYPE_Q8_0:
+                        {
+                            switch (dstt) {
+                                case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F32].pipeline; break;
+                                case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F16].pipeline; break;
+                                default: GGML_ABORT("not implemented");
                            };
                        } break;
                    default: GGML_ABORT("not implemented");
@@ -3966,7 +4034,11 @@ static void ggml_metal_encode_node(
                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];

+                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+                int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
+
                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+
            } break;
        case GGML_OP_SET:
            {
@@ -4341,6 +4341,49 @@ kernel void kernel_cpy_f32_iq4_nl(
    }
 }

+template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_cpy_q_f32(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int64_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T4x4    * dst_data = (device       T4x4    *)(dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1 + i0*args.nb0);
+
+    for (int64_t i00 = tpitg.x; i00 < args.ne00/16; i00 += ntg.x) {
+        T4x4 temp;
+        dequantize_func(src_data + i00/nl, i00%nl, temp);
+        dst_data[i00] = temp;
+    }
+}
+
+typedef decltype(kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>) cpy_q_f_t;
+
+template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;
+
 kernel void kernel_concat(
    constant ggml_metal_kargs_concat & args,
    device  const char * src0,
@@ -83,6 +83,10 @@ if (MUSAToolkit_FOUND)
        add_compile_definitions(GGML_CUDA_NO_VMM)
    endif()

+    if (NOT GGML_CUDA_FA)
+        add_compile_definitions(GGML_CUDA_NO_FA)
+    endif()
+
    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
        add_compile_definitions(GGML_CUDA_F16)
    endif()
@@ -444,19 +444,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
        backend_ctx->gpu_family = GPU_FAMILY::ADRENO;
        backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->name);

-        // Default wave size is 128, A8x uses 64.
-        if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::A8X) {
-            backend_ctx->adreno_wave_size = 64;
-        } else if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X ||
-                   backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E) {
-            backend_ctx->adreno_wave_size = 128;
-        } else {
-            backend_ctx->adreno_wave_size = 128;
-            GGML_LOG_WARN("ggml_opencl: Unsupported Adreno GPU: %s, "
-                "using wave size %d, "
-                "may not work as expected\n",
-                backend_ctx->device_name.c_str(), backend_ctx->adreno_wave_size);
-        }
+        // Use wave size of 64 for all Adreno GPUs.
+        backend_ctx->adreno_wave_size = 64;
    } else if (strstr(default_device->name, "Intel")) {
        backend_ctx->gpu_family = GPU_FAMILY::INTEL;
    } else {
@@ -1376,6 +1365,11 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        int M = tensor->ne[1];   // ne01
        int K = tensor->ne[0];   // ne00

+        //For matrix-vector multiplication kernel, we assume K is a multiple of 32
+        GGML_ASSERT(K % 32 == 0);
+        //For transpose kernels, we assume K is a multiple of 4 (satisfied by prior assert), and M is a multiple of 4
+        GGML_ASSERT(M % 4 == 0);
+
        // transpose is out of place, so we need to allocate transposed buffers
        // <----------------------------------------------------------------------------------> //
        // use sub_buffer of max buffer size instead
@@ -1416,36 +1410,36 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        cl_mem qT_d_image1D;
        cl_mem dT_d_image1D;

-        cl_image_format img_fmt_1d = { CL_RGBA, CL_FLOAT };
+        cl_image_format img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
        cl_image_desc img_desc_1d;

        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 8 / 4;
+        img_desc_1d.image_width = M * K / 4 / 4;
        img_desc_1d.buffer = extra->q;
        q_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
        CL_CHECK(err);

-        img_fmt_1d = { CL_RGBA, CL_FLOAT };
+        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 8 / 4;
+        img_desc_1d.image_width = M * K / 4 / 4;
        img_desc_1d.buffer = qT_d;
        qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
        CL_CHECK(err);

-        img_fmt_1d = { CL_RGBA, CL_FLOAT };
+        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 32 / 4 / 2;
+        img_desc_1d.image_width = M * K / 32 / 4;
        img_desc_1d.buffer = extra->d;
        d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
        CL_CHECK(err);

-        img_fmt_1d = { CL_RGBA, CL_FLOAT };
+        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 32 / 4 / 2;
+        img_desc_1d.image_width = M * K / 32 / 4;
        img_desc_1d.buffer = dT_d;
        dT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
        CL_CHECK(err);
@@ -1454,8 +1448,8 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        // set up and call the transpose kernels
        // <----------------------------------------------------------------------------------> //
        // weights
-        int height_q = M / 8;
-        int width_q = K / 8 / 4;
+        int height_q = M / 4;
+        int width_q = K / 4 / 4;
        kernel = backend_ctx->kernel_transpose_16;

        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &q_d_image1D));
@@ -1469,8 +1463,8 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        CL_CHECK(clWaitForEvents(1, &evt));

        // scales
-        int height_s = M / 8;
-        int width_s = K / 32 / 8;
+        int height_s = M / 4;
+        int width_s = K / 32 / 4;

        kernel = backend_ctx->kernel_transpose_16;
        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
@@ -1864,7 +1858,6 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
    void * buf_d;
 #endif

-#ifdef GGML_USE_OPENCL
    // Make sure everything is done.
    CL_CHECK(clFinish(queue));

@@ -1900,7 +1893,6 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
        extra->offset, ggml_nbytes(tensor), buf, 0, NULL, NULL));
    CL_CHECK(clFinish(queue));
 #endif // GGML_OPENCL_SOA_Q
-#endif // GGML_USE_OPENCL

    // Open file and dump.
    char fname[512];
@@ -2865,6 +2857,9 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            CL_CHECK(status);

            int height_B = N/4;
+            if (height_B == 0) {
+                height_B = 1;
+            }
            int width_B = K/4;
            int padded_height_B = (N + padding)/4;

@@ -3013,11 +3008,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
        }

        if (N == 1) {
-            local_work_size[0] = backend_ctx->adreno_wave_size; // localsize
+            size_t wavesize = backend_ctx->adreno_wave_size;
+            local_work_size[0] = wavesize; // localsize
            local_work_size[1] = 4; // reduce factor
            local_work_size[2] = 1;

-            global_work_size[0] = M / 2;
+            global_work_size[0] = (((M / 2) + wavesize - 1) / wavesize) * wavesize;
            global_work_size[1] = 4; // reduce factor
            global_work_size[2] = 1;
        }
@@ -1797,6 +1797,9 @@ kernel void kernel_mul_mat_f16_f16(
 //------------------------------------------------------------------------------
 // mul_mat_f16_f32_1row
 //------------------------------------------------------------------------------
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
 kernel void kernel_mul_mat_f16_f32_1row(
        global char * src0,
        ulong offset0,
@@ -1,9 +1,11 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#pragma OPENCL EXTENSION cl_qcom_subgroup_uniform_load: enable
-#pragma OPENCL EXTENSION cl_qcom_subgroup_constant_load: enable
-#pragma OPENCL EXTENSION cl_qcom_extra_vector_types : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif

 // assume
 #define QK4_0 32
@@ -186,8 +188,9 @@
    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \

-
-__attribute__((qcom_reqd_sub_group_size("full")))
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
 __kernel void kernel_gemv_noshuffle(
        __read_only  image1d_buffer_t src0_q,  // quantized A
        global half2  * src0_d,  // A scales
@@ -1,9 +1,11 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #pragma OPENCL EXTENSION cl_khr_subgroups : enable
-#pragma OPENCL EXTENSION cl_qcom_subgroup_uniform_load: enable
-#pragma OPENCL EXTENSION cl_qcom_subgroup_constant_load: enable
-#pragma OPENCL EXTENSION cl_qcom_extra_vector_types : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif

 // assume
 #define QK4_0 32
@@ -186,8 +188,9 @@
    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \

-
-__attribute__((qcom_reqd_sub_group_size("full")))
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
 __kernel void kernel_gemv_noshuffle(
        __read_only  image1d_buffer_t src0_q,  // quantized A
        global half2  * src0_d,  // A scales
@@ -7,7 +7,16 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable

-__attribute__((qcom_reqd_sub_group_size("full")))
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
 kernel void kernel_mul_mat_Ab_Bi_8x4(
        global const ushort * src0_q,       // quantized A
        global const half  * src0_d,        // A scales
@@ -1,4 +1,6 @@
-// 16-bit transpose, loading/storing an 8x8 tile of elements
+// 16-bit transpose, loading/storing a 4x4 tile of elements
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable

 kernel void kernel_transpose_16(
    __read_only image1d_buffer_t input,
@@ -9,24 +11,16 @@ kernel void kernel_transpose_16(

    const int i = get_global_id(0);
    const int j = get_global_id(1);
-    const int i_3 = i<<3;
-    const int j_3 = j<<3;
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;

-    ushort8 temp0 = as_ushort8(read_imagef(input, (j_3+0)*cols+i));
-    ushort8 temp1 = as_ushort8(read_imagef(input, (j_3+1)*cols+i));
-    ushort8 temp2 = as_ushort8(read_imagef(input, (j_3+2)*cols+i));
-    ushort8 temp3 = as_ushort8(read_imagef(input, (j_3+3)*cols+i));
-    ushort8 temp4 = as_ushort8(read_imagef(input, (j_3+4)*cols+i));
-    ushort8 temp5 = as_ushort8(read_imagef(input, (j_3+5)*cols+i));
-    ushort8 temp6 = as_ushort8(read_imagef(input, (j_3+6)*cols+i));
-    ushort8 temp7 = as_ushort8(read_imagef(input, (j_3+7)*cols+i));
+    half4 temp0 = read_imageh(input, (j_2+0)*cols+i);
+    half4 temp1 = read_imageh(input, (j_2+1)*cols+i);
+    half4 temp2 = read_imageh(input, (j_2+2)*cols+i);
+    half4 temp3 = read_imageh(input, (j_2+3)*cols+i);

-    write_imagef(output, (i_3+0)*rows+j, as_float4((ushort8)(temp0.s0, temp1.s0, temp2.s0, temp3.s0, temp4.s0, temp5.s0, temp6.s0, temp7.s0)));
-    write_imagef(output, (i_3+1)*rows+j, as_float4((ushort8)(temp0.s1, temp1.s1, temp2.s1, temp3.s1, temp4.s1, temp5.s1, temp6.s1, temp7.s1)));
-    write_imagef(output, (i_3+2)*rows+j, as_float4((ushort8)(temp0.s2, temp1.s2, temp2.s2, temp3.s2, temp4.s2, temp5.s2, temp6.s2, temp7.s2)));
-    write_imagef(output, (i_3+3)*rows+j, as_float4((ushort8)(temp0.s3, temp1.s3, temp2.s3, temp3.s3, temp4.s3, temp5.s3, temp6.s3, temp7.s3)));
-    write_imagef(output, (i_3+4)*rows+j, as_float4((ushort8)(temp0.s4, temp1.s4, temp2.s4, temp3.s4, temp4.s4, temp5.s4, temp6.s4, temp7.s4)));
-    write_imagef(output, (i_3+5)*rows+j, as_float4((ushort8)(temp0.s5, temp1.s5, temp2.s5, temp3.s5, temp4.s5, temp5.s5, temp6.s5, temp7.s5)));
-    write_imagef(output, (i_3+6)*rows+j, as_float4((ushort8)(temp0.s6, temp1.s6, temp2.s6, temp3.s6, temp4.s6, temp5.s6, temp6.s6, temp7.s6)));
-    write_imagef(output, (i_3+7)*rows+j, as_float4((ushort8)(temp0.s7, temp1.s7, temp2.s7, temp3.s7, temp4.s7, temp5.s7, temp6.s7, temp7.s7)));
+    write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
 }
@@ -1,3 +1,5 @@
+message(STATUS  "GGML_SYCL_TARGET=${GGML_SYCL_TARGET}")
+
 if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
    message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
 endif()
@@ -99,3 +99,20 @@ catch (sycl::exception const &exc) {
            << ", line:" << __LINE__ << std::endl;
  std::exit(1);
 }
+
+
+void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> streams) {
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) {
+            if (extra->events[i][is] != nullptr) {
+                SYCL_CHECK(CHECK_TRY_ERROR(dpct::destroy_event(extra->events[i][is])));
+            }
+        }
+        if (extra->data_device[i] != nullptr && streams.size()>0) {
+            ggml_sycl_set_device(i);
+            SYCL_CHECK(
+                CHECK_TRY_ERROR(sycl::free(extra->data_device[i], *(streams[i]))));
+        }
+    }
+    delete extra;
+}
@@ -19,6 +19,9 @@
 #include "dpct/helper.hpp"
 #include "ggml-sycl.h"
 #include "presets.hpp"
+#include "sycl_hw.hpp"
+
+
 #if GGML_SYCL_DNNL
 #include "dnnl.hpp"
 #include "dnnl_sycl.hpp"
@@ -35,7 +38,10 @@
 void* ggml_sycl_host_malloc(size_t size);
 void ggml_sycl_host_free(void* ptr);

-static int g_ggml_sycl_debug = 0;
+
+extern int g_ggml_sycl_debug;
+extern int g_ggml_sycl_disable_optimize;
+
 #define GGML_SYCL_DEBUG(...)        \
  do {                              \
    if (g_ggml_sycl_debug)          \
@@ -182,18 +188,24 @@ inline dpct::err0 ggml_sycl_set_device(const int device) try {
 }

 //////////////////////
+struct optimize_feature {
+    bool reorder=false;
+};
+
+struct sycl_device_info {
+    int     cc;                 // compute capability
+    // int     nsm;                // number of streaming multiprocessors
+    // size_t  smpb;               // max. shared memory per block
+    bool    vmm;                // virtual memory support
+    size_t  total_vram;
+    sycl_hw_info hw_info;
+    optimize_feature opt_feature;
+};
+

 struct ggml_sycl_device_info {
    int device_count;

-    struct sycl_device_info {
-        int     cc;                 // compute capability
-        // int     nsm;                // number of streaming multiprocessors
-        // size_t  smpb;               // max. shared memory per block
-        bool    vmm;                // virtual memory support
-        size_t  total_vram;
-    };
-
    sycl_device_info devices[GGML_SYCL_MAX_DEVICES] = {};

    std::array<float, GGML_SYCL_MAX_DEVICES> default_tensor_split = {};
@@ -260,17 +272,46 @@ struct ggml_tensor_extra_gpu {
                                       // tensors
  dpct::event_ptr events[GGML_SYCL_MAX_DEVICES]
                        [GGML_SYCL_MAX_STREAMS]; // events for synchronizing multiple GPUs
+  optimize_feature optimized_feature;
 };

+void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> streams={});
+
+inline optimize_feature check_gpu_optimize_feature(syclex::architecture &arch) {
+    optimize_feature opt;
+
+    opt.reorder =
+        (arch == syclex::architecture::intel_gpu_dg1 ||
+         arch == syclex::architecture::intel_gpu_acm_g10 ||
+         arch == syclex::architecture::intel_gpu_acm_g11 ||
+         arch == syclex::architecture::intel_gpu_acm_g12 ||
+         arch == syclex::architecture::intel_gpu_pvc ||
+         arch == syclex::architecture::intel_gpu_pvc_vg ||
+         arch == syclex::architecture::intel_gpu_mtl_u ||
+         arch == syclex::architecture::intel_gpu_mtl_s ||
+         arch == syclex::architecture::intel_gpu_mtl_h ||
+         arch == syclex::architecture::intel_gpu_arl_u ||
+         arch == syclex::architecture::intel_gpu_arl_s ||
+         arch == syclex::architecture::intel_gpu_arl_h ||
+         arch == syclex::architecture::intel_gpu_bmg_g21 ||
+         arch == syclex::architecture::intel_gpu_lnl_m
+        );
+
+    return opt;
+}
+
 struct ggml_backend_sycl_context {
    int device;
    std::string name;
+    optimize_feature opt_feature;
+    bool optimized_graph=false;

    queue_ptr qptrs[GGML_SYCL_MAX_DEVICES][GGML_SYCL_MAX_STREAMS] = { { nullptr } };

    explicit ggml_backend_sycl_context(int device) :
        device(device),
        name(GGML_SYCL_NAME + std::to_string(device)) {
+        opt_feature = ggml_sycl_info().devices[device].opt_feature;
    }

    queue_ptr stream(int device, int stream) {
@@ -680,5 +721,4 @@ bool gpu_has_xmx(sycl::device &dev);
 void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                 const ggml_tensor *src1, ggml_tensor *dst,
                                 const ggml_sycl_op_flatten_t op);
-
 #endif // GGML_SYCL_COMMON_HPP
@@ -125,6 +125,25 @@ static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k,
    }
 }

+template <typename dst_t>
+static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+
+    dpct::has_capability_or_fail(stream->get_device(),
+                                    {sycl::aspect::fp16});
+
+    int constexpr WARP_K = WARP_SIZE * QK4_0;
+    const int n_warp = (k + WARP_K - 1) / WARP_K;
+    GGML_ASSERT(k % 2 == 0);
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) *
+        sycl::range<3>(1, 1, WARP_SIZE),
+        sycl::range<3>(1, 1, WARP_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]]{
+            dequantize_block_q4_0_reorder(vx, y, k, item_ct1);
+        });
+
+}
+
 template <typename dst_t>
 static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k,
                                     dpct::queue_ptr stream) {
@@ -452,10 +471,15 @@ static void convert_unary_sycl(const void *__restrict__ vx,
    }
 }

-to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) {
+to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor *dst) {
    switch (type) {
        case GGML_TYPE_Q4_0:
-            return dequantize_block_sycl<QK4_0, QR4_0, dequantize_q4_0>;
+            if (dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q4_0_sycl_reorder;
+            } else {
+                return dequantize_block_sycl<QK4_0, QR4_0, dequantize_q4_0>;
+            }
        case GGML_TYPE_Q4_1:
            return dequantize_block_sycl<QK4_1, QR4_1, dequantize_q4_1>;
        case GGML_TYPE_Q5_0:
@@ -499,10 +523,15 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) {
    }
 }

-to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) {
+to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
    switch (type) {
        case GGML_TYPE_Q4_0:
-            return dequantize_row_q4_0_sycl;
+            if (dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q4_0_sycl_reorder;
+            } else {
+                return dequantize_row_q4_0_sycl;
+            }
        case GGML_TYPE_Q4_1:
            return dequantize_row_q4_1_sycl;
        case GGML_TYPE_Q5_0:
@@ -21,7 +21,7 @@ using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y,
 typedef to_t_sycl_t<float> to_fp32_sycl_t;
 typedef to_t_sycl_t<sycl::half> to_fp16_sycl_t;

-to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type);
-to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type);
+to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor *dst);
+to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst);

 #endif // GGML_SYCL_CONVERT_HPP
@@ -16,6 +16,8 @@
 #include "common.hpp"

 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
+typedef void (*dequantize_kernel_t_reorder)(const void *d, const int64_t ib, const void *qs,
+                                            const int iqs, dfloat2 &v);

 static __dpct_inline__ void dequantize_q4_0(const void *vx, const int64_t ib,
                                            const int iqs, dfloat2 &v) {
@@ -40,6 +42,29 @@ static __dpct_inline__ void dequantize_q4_0(const void *vx, const int64_t ib,
 #endif // GGML_SYCL_F16
 }

+static __dpct_inline__ void dequantize_q4_0_reorder(const void *d_ptr, const int64_t ib, const void *qs,
+                                            const int iqs, dfloat2 &v) {
+    // const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const dfloat d = (const dfloat)*((const sycl::half*)d_ptr+ib);
+
+    const int vui = *((const uint8_t *)qs+iqs);
+
+    v.x() = vui & 0xF;
+    v.y() = vui >> 4;
+
+#ifdef GGML_SYCL_F16
+    // v = v - {8.0f, 8.0f};
+    // v = v * {d, d};
+    v.s0() = (v.s0() - 8.0f) * d;
+    v.s1() = (v.s1() - 8.0f) * d;
+
+#else
+    v.x() = (v.x() - 8.0f) * d;
+    v.y() = (v.y() - 8.0f) * d;
+#endif // GGML_SYCL_F16
+}
+
 static __dpct_inline__ void dequantize_q4_1(const void *vx, const int64_t ib,
                                            const int iqs, dfloat2 &v) {
    const block_q4_1 * x = (const block_q4_1 *) vx;
@@ -167,6 +192,36 @@ static void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restri
    }
 }

+template<typename dst_t>
+static void dequantize_block_q4_0_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t nb32,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+    auto k=nb32;
+    // assume 32 threads
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int lane_ib = i * WARP_SIZE + tid;
+
+    if (lane_ib >= k / QK4_0) {
+        return;
+    }
+
+    dst_t * y_ptr = yy + lane_ib * QK4_0;
+
+    auto qs = (const uint8_t*)vx + lane_ib * QK4_0 / 2;
+    auto s_ptr = (const sycl::half*)((const uint8_t*)vx + k / 2) + lane_ib;
+
+    const float d = float(*s_ptr);
+
+#pragma unroll
+    for (int l = 0; l < QK4_0 / 2; ++l) {
+        int vq = qs[l];
+        y_ptr[l + 0] = d * ((vq & 0xF) - 8);
+        y_ptr[l + 16] = d * ((vq >> 4) - 8);
+    }
+
+}
+
 template<typename dst_t>
 static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t nb32,
                                  const sycl::nd_item<3> &item_ct1) {
@@ -3,7 +3,6 @@
 #include "dequantize.hpp"
 #include "presets.hpp"

-
 static void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
    const sycl::half *x = (const sycl::half *)vx;

@@ -91,6 +90,112 @@ static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat *
    }
 }

+template <int qk, int qr, dequantize_kernel_t_reorder dequantize_kernel_reorder>
+static void dequantize_mul_mat_vec_reorder(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows,
+                                   const sycl::nd_item<3> &item_ct1) {
+    // qk = quantized weights per x block
+    // qr = number of quantized weights per data value in x block
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int tid = item_ct1.get_local_id(2);
+
+
+    const int ncols_left = ncols % (QK4_0*WARP_SIZE);
+    const int ncols_align = ncols - ncols_left;
+    const int iter_stride = 8*2*GGML_SYCL_DMMV_X;
+    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter //64/16=4, 512/16/2= 16
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+// partial sum for each thread
+#ifdef GGML_SYCL_F16
+    sycl::half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
+#else
+    float tmp = 0.0f;
+#endif // GGML_SYCL_F16
+    const char *d_ptr = (const char*)vx+ncols*nrows/2;
+    int i=0;
+    for (i = 0; i < ncols_align; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int ib = (row*ncols + col)/qk; // x block index
+        const int iqs = (col%qk)/qr; // x quant index
+        const int iybs = col - col%qk; // y block start index
+
+// processing >2 values per i iter is faster for fast GPUs
+#pragma unroll
+        for (int j = 0; j < vals_per_iter; j += 2) {
+            // process 2 vals per j iter
+
+            // dequantize
+            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
+            dfloat2 v;
+            dequantize_kernel_reorder((const void *)d_ptr, ib, (const void *)vx, ib * QK4_0 / 2 +iqs+j/qr, v);
+
+            // matrix multiplication
+            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+#ifdef GGML_SYCL_F16
+            dfloat2 t1{y[iybs + iqs + j / qr + 0],
+                        y[iybs + iqs + j / qr + y_offset]};
+
+            tmp += v * t1;
+#else
+            tmp += v.x() * y[iybs + iqs + j / qr + 0];
+            tmp += v.y() * y[iybs + iqs + j / qr + y_offset];
+#endif // GGML_SYCL_F16
+        }
+    }
+
+    for (; i < ncols; i += iter_stride) {
+        if (tid>=ncols_left/QK4_0) continue;
+        const int col = i + vals_per_iter*tid;
+        const int ib = (row*ncols + col)/qk; // x block index
+        const int iqs = (col%qk)/qr; // x quant index
+        const int iybs = col - col%qk; // y block start index
+
+// processing >2 values per i iter is faster for fast GPUs
+#pragma unroll
+        for (int j = 0; j < vals_per_iter; j += 2) {
+            // process 2 vals per j iter
+
+            // dequantize
+            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
+            dfloat2 v;
+            dequantize_kernel_reorder((const void *)d_ptr, ib, (const void *)vx, ib * QK4_0 / 2 +iqs+j/qr, v);
+
+            // matrix multiplication
+            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+#ifdef GGML_SYCL_F16
+            dfloat2 t1{y[iybs + iqs + j / qr + 0],
+                        y[iybs + iqs + j / qr + y_offset]};
+
+            tmp += v * t1;
+#else
+            tmp += v.x() * y[iybs + iqs + j / qr + 0];
+            tmp += v.y() * y[iybs + iqs + j / qr + y_offset];
+#endif // GGML_SYCL_F16
+        }
+    }
+
+    // sum up partial sums and write back result
+    const int mask_start = ncols > GGML_SYCL_DMMV_X ? WARP_SIZE >> 1 : WARP_SIZE >> 2;
+    for (int mask = mask_start; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (tid == 0) {
+#ifdef GGML_SYCL_F16
+        dst[row] = tmp.x() + tmp.y();
+#else
+        dst[row] = tmp;
+#endif // GGML_SYCL_F16
+    }
+}
+
 static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
                                         float *dst, const int ncols,
                                         const int nrows,
@@ -759,6 +864,28 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa
    }
 }

+static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloat *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
+    }
+}
+

 static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y,
                                             float *dst, const int ncols,
@@ -953,7 +1080,6 @@ void ggml_sycl_op_dequantize_mul_mat_vec(

    const int64_t ne00 = src0->ne[0];
    const int64_t row_diff = row_high - row_low;
-
    GGML_ASSERT(src1->type == GGML_TYPE_F32);
    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
 #ifdef GGML_SYCL_F16
@@ -967,7 +1093,7 @@ void ggml_sycl_op_dequantize_mul_mat_vec(

    if (src1_convert_f16) {
        src1_dfloat = src1_dfloat_a.alloc(ne00);
-        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type);
+        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
        GGML_ASSERT(to_fp16_sycl != nullptr);
        to_fp16_sycl(src1_ddf_i, src1_dfloat, ne00, stream);
    }
@@ -977,7 +1103,12 @@ void ggml_sycl_op_dequantize_mul_mat_vec(

    switch (src0->type) {
        case GGML_TYPE_Q4_0:
-            dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            if ((ggml_tensor_extra_gpu*)dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
+                dequantize_mul_mat_vec_q4_0_sycl_reorder(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            } else {
+                dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            }
            break;
        case GGML_TYPE_Q4_1:
            dequantize_mul_mat_vec_q4_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
@@ -1020,4 +1151,5 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
    GGML_UNUSED(src1_ddq_i);
    GGML_UNUSED(src1_ncols);
    GGML_UNUSED(src1_padded_row_size);
+    GGML_UNUSED(ctx);
 }
--- a/Show More
+++ b/Show More