vulkan: workaround for AMD Windows driver 16 bit unpack8 bug (#12472)

ggml-ci

.github/workflows/build.yml

@@ -676,6 +676,35 @@ jobs:
             -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
+  macOS-latest-cmake-visionos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=visionOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
   macOS-latest-swift:
     runs-on: macos-latest
 

build-xcframework.sh

@@ -432,8 +432,8 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
     -S .
 cmake --build build-visionos --config Release -- -quiet
 
@@ -445,8 +445,8 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 -Du_int=unsigned\ int -Du_char=unsigned\ char -Du_short=unsigned\ short ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet
 

convert_hf_to_gguf.py

@@ -180,7 +180,8 @@ class Model:
             extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
             missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
             if len(extra) == 0 and len(missing_files) > 0:
-                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+                raise ValueError(f"Missing or incomplete model files: {missing_files}\n"
+                                 f"Missing tensors: {missing}")
             else:
                 raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
                                  f"Missing tensors: {missing}\n"
@@ -528,6 +529,8 @@ class Model:
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
         added_vocab = tokenizer.get_added_vocab()
 
+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
         for i in range(vocab_size):
             if i not in reverse_vocab:
                 tokens.append(f"[PAD{i}]")
@@ -537,13 +540,13 @@ class Model:
                 if token in added_vocab:
                     # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
                     # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not tokenizer.added_tokens_decoder[i].normalized:
+                    if not added_tokens_decoder[i].normalized:
                         previous_token = token
                         token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
                         if previous_token != token:
                             logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
 
-                    if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
                         toktypes.append(gguf.TokenType.CONTROL)
                     else:
                         # NOTE: this was added for Gemma.
@@ -1099,13 +1102,6 @@ class BloomModel(Model):
 
         tensors.append((self.map_tensor_name(name), data_torch))
 
-        if name == "word_embeddings.weight":
-            assert self.tensor_names is not None
-
-            # TODO: tie them at runtime, don't duplicate in the model file
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -2423,10 +2419,6 @@ class GPT2Model(Model):
 
         tensors.append((new_name, data_torch))
 
-        # note: GPT2 output is tied to (same as) wte in original model
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
         return tensors
 
 
@@ -2756,21 +2748,26 @@ class CodeShellModel(Model):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
         self.gguf_writer.add_rope_scaling_factor(1.0)
 
+    _has_tok_embd = False
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
+        output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
+        tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD)
+
         new_name = self.map_tensor_name(name)
 
-        tensors: list[tuple[str, Tensor]] = [(new_name, data_torch)]
+        # assuming token_embd.weight is seen before output.weight
+        if not self._has_tok_embd and new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
+            # even though the tensor file(s) does not contain the word embeddings they are still in the weight map
+            if self.tensor_names and "transformer.wte.weight" in self.tensor_names:
+                logger.debug(f"{tok_embd_name} not found before {output_name}, assuming they are tied")
+                self.tensor_names.remove("transformer.wte.weight")
+        elif new_name == tok_embd_name:
+            self._has_tok_embd = True
 
-        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            assert self.tensor_names is not None
-
-            if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")):
-                # copy tok_embd.weight to output.weight
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch))
-
-        return tensors
+        return [(new_name, data_torch)]
 
 
 @Model.register("InternLM2ForCausalLM")

docs/backend/SYCL.md

@@ -237,6 +237,15 @@ cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENAB
 cmake --build buildWithCublas --config Release
 ```
 
+**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
+
+```sh
+git clone https://github.com/oneapi-src/oneDNN.git
+cd oneDNN
+cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake --build build-nvidia --config Release
+```
+
 - **Adding support to AMD GPUs**
 
 **oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
@@ -327,10 +336,10 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
 
 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
 
 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
 
 # build all binary
 cmake --build build --config Release -j -v

examples/server/webui/src/components/ChatScreen.tsx

@@ -99,13 +99,9 @@ export default function ChatScreen() {
     canvasData,
     replaceMessageAndGenerate,
   } = useAppContext();
-  const [inputMsg, setInputMsg] = useState(prefilledMsg.content());
-  const inputRef = useRef<HTMLTextAreaElement>(null);
+  const textarea = useOptimizedTextarea(prefilledMsg.content());
 
-  const { extraContext, clearExtraContext } = useVSCodeContext(
-    inputRef,
-    setInputMsg
-  );
+  const { extraContext, clearExtraContext } = useVSCodeContext(textarea);
   // TODO: improve this when we have "upload file" feature
   const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined;
 
@@ -135,9 +131,10 @@ export default function ChatScreen() {
   };
 
   const sendNewMessage = async () => {
-    if (inputMsg.trim().length === 0 || isGenerating(currConvId ?? '')) return;
-    const lastInpMsg = inputMsg;
-    setInputMsg('');
+    const lastInpMsg = textarea.value();
+    if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? ''))
+      return;
+    textarea.setValue('');
     scrollToBottom(false);
     setCurrNodeId(-1);
     // get the last message node
@@ -146,13 +143,13 @@ export default function ChatScreen() {
       !(await sendMessage(
         currConvId,
         lastMsgNodeId,
-        inputMsg,
+        lastInpMsg,
         currExtra,
         onChunk
       ))
     ) {
       // restore the input message if failed
-      setInputMsg(lastInpMsg);
+      textarea.setValue(lastInpMsg);
     }
     // OK
     clearExtraContext();
@@ -195,16 +192,13 @@ export default function ChatScreen() {
       // send the prefilled message if needed
       sendNewMessage();
     } else {
-      // otherwise, focus on the input and move the cursor to the end
-      if (inputRef.current) {
-        inputRef.current.focus();
-        inputRef.current.selectionStart = inputRef.current.value.length;
-      }
+      // otherwise, focus on the input
+      textarea.focus();
     }
     prefilledMsg.clear();
     // no need to keep track of sendNewMessage
     // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [inputRef]);
+  }, [textarea.ref]);
 
   // due to some timing issues of StorageUtils.appendMsg(), we need to make sure the pendingMsg is not duplicated upon rendering (i.e. appears once in the saved conversation and once in the pendingMsg)
   const pendingMsgDisplay: MessageDisplay[] =
@@ -258,9 +252,7 @@ export default function ChatScreen() {
           <textarea
             className="textarea textarea-bordered w-full"
             placeholder="Type a message (Shift+Enter to add a new line)"
-            ref={inputRef}
-            value={inputMsg}
-            onChange={(e) => setInputMsg(e.target.value)}
+            ref={textarea.ref}
             onKeyDown={(e) => {
               if (e.nativeEvent.isComposing || e.keyCode === 229) return;
               if (e.key === 'Enter' && e.shiftKey) return;
@@ -280,11 +272,7 @@ export default function ChatScreen() {
               Stop
             </button>
           ) : (
-            <button
-              className="btn btn-primary ml-2"
-              onClick={sendNewMessage}
-              disabled={inputMsg.trim().length === 0}
-            >
+            <button className="btn btn-primary ml-2" onClick={sendNewMessage}>
               Send
             </button>
           )}
@@ -298,3 +286,43 @@ export default function ChatScreen() {
     </div>
   );
 }
+
+export interface OptimizedTextareaValue {
+  value: () => string;
+  setValue: (value: string) => void;
+  focus: () => void;
+  ref: React.RefObject<HTMLTextAreaElement>;
+}
+
+// This is a workaround to prevent the textarea from re-rendering when the inner content changes
+// See https://github.com/ggml-org/llama.cpp/pull/12299
+function useOptimizedTextarea(initValue: string): OptimizedTextareaValue {
+  const [savedInitValue, setSavedInitValue] = useState<string>(initValue);
+  const textareaRef = useRef<HTMLTextAreaElement>(null);
+
+  useEffect(() => {
+    if (textareaRef.current && savedInitValue) {
+      textareaRef.current.value = savedInitValue;
+      setSavedInitValue('');
+    }
+  }, [textareaRef, savedInitValue, setSavedInitValue]);
+
+  return {
+    value: () => {
+      return textareaRef.current?.value ?? savedInitValue;
+    },
+    setValue: (value: string) => {
+      if (textareaRef.current) {
+        textareaRef.current.value = value;
+      }
+    },
+    focus: () => {
+      if (textareaRef.current) {
+        // focus and move the cursor to the end
+        textareaRef.current.focus();
+        textareaRef.current.selectionStart = textareaRef.current.value.length;
+      }
+    },
+    ref: textareaRef,
+  };
+}

examples/server/webui/src/utils/llama-vscode.ts

@@ -1,5 +1,6 @@
 import { useEffect, useState } from 'react';
 import { MessageExtraContext } from './types';
+import { OptimizedTextareaValue } from '../components/ChatScreen';
 
 // Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe
 // Ref: https://github.com/ggml-org/llama.cpp/pull/11940
@@ -14,10 +15,7 @@ interface SetTextEvData {
  * window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n  return 123' }, '*');
  */
 
-export const useVSCodeContext = (
-  inputRef: React.RefObject<HTMLTextAreaElement>,
-  setInputMsg: (text: string) => void
-) => {
+export const useVSCodeContext = (textarea: OptimizedTextareaValue) => {
   const [extraContext, setExtraContext] = useState<MessageExtraContext | null>(
     null
   );
@@ -27,20 +25,20 @@ export const useVSCodeContext = (
     const handleMessage = (event: MessageEvent) => {
       if (event.data?.command === 'setText') {
         const data: SetTextEvData = event.data;
-        setInputMsg(data?.text);
+        textarea.setValue(data?.text);
         if (data?.context && data.context.length > 0) {
           setExtraContext({
             type: 'context',
             content: data.context,
           });
         }
-        inputRef.current?.focus();
+        textarea.focus();
       }
     };
 
     window.addEventListener('message', handleMessage);
     return () => window.removeEventListener('message', handleMessage);
-  }, [inputRef, setInputMsg]);
+  }, [textarea]);
 
   // Add a keydown listener that sends the "escapePressed" message to the parent window
   useEffect(() => {

examples/tts/tts.cpp

@@ -571,6 +571,10 @@ int main(int argc, char ** argv) {
     model_ttc = llama_init_ttc.model.get();
     ctx_ttc   = llama_init_ttc.context.get();
 
+    if (model_ttc == nullptr || ctx_ttc == nullptr) {
+        return ENOENT;
+    }
+
     const llama_vocab * vocab = llama_model_get_vocab(model_ttc);
 
     // TODO: refactor in a common struct
@@ -586,6 +590,10 @@ int main(int argc, char ** argv) {
     model_cts = llama_init_cts.model.get();
     ctx_cts   = llama_init_cts.context.get();
 
+    if (model_cts == nullptr || ctx_cts == nullptr) {
+        return ENOENT;
+    }
+
     std::vector<common_sampler *> smpl(n_parallel);
     for (int i = 0; i < n_parallel; ++i) {
         params.sampling.no_perf = (i != 0);

ggml/src/CMakeLists.txt

@@ -76,7 +76,11 @@ if (GGML_CCACHE)
             set(GGML_CCACHE_VARIANT sccache)
         endif()
         # TODO: should not be set globally
-        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
+        if (GGML_SYCL AND GGML_CCACHE_FOUND AND WIN32)
+            set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "ccache compiler_type=icl")
+        else ()
+            set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}")
+        endif ()
         set(ENV{CCACHE_SLOPPINESS} time_macros)
         message(STATUS "${GGML_CCACHE_VARIANT} found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
     else()
@@ -325,6 +329,10 @@ if (CMAKE_SYSTEM_NAME MATCHES "Android")
     target_link_libraries(ggml-base PRIVATE dl)
 endif()
 
+if(CMAKE_SYSTEM_NAME MATCHES "visionOS")
+    target_compile_definitions(ggml-base PUBLIC _DARWIN_C_SOURCE)
+endif()
+
 if (BUILD_SHARED_LIBS)
     foreach (target ggml-base ggml)
         set_target_properties(${target} PROPERTIES POSITION_INDEPENDENT_CODE ON)

ggml/src/ggml-cuda/fattn-common.cuh

@@ -606,48 +606,47 @@ static __global__ void flash_attn_stream_k_fixup(
     *dst = dst_val / rowsum;
 }
 
-template<int D, int parallel_blocks> // D == head size
+template<int D> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_combine_results(
         const float  * __restrict__ VKQ_parts,
         const float2 * __restrict__ VKQ_meta,
-        float * __restrict__ dst) {
-    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
-    dst       +=                 D * gridDim.y*blockIdx.x;
+        float * __restrict__ dst,
+        const int parallel_blocks) {
+    VKQ_parts += parallel_blocks*D * gridDim.z*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.z*blockIdx.x;
+    dst       +=                 D * gridDim.z*blockIdx.x;
 
     const int tid = threadIdx.x;
     __builtin_assume(tid < D);
 
-    __shared__ float2 meta[parallel_blocks];
+    extern __shared__ float2 meta[];
     if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + tid];
     }
 
     __syncthreads();
 
     float kqmax = meta[0].x;
-#pragma unroll
     for (int l = 1; l < parallel_blocks; ++l) {
         kqmax = max(kqmax, meta[l].x);
     }
 
     float VKQ_numerator   = 0.0f;
     float VKQ_denominator = 0.0f;
-#pragma unroll
     for (int l = 0; l < parallel_blocks; ++l) {
         const float diff = meta[l].x - kqmax;
         const float KQ_max_scale = expf(diff);
         const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
         *((uint32_t *) &KQ_max_scale) &= ftz_mask;
 
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.z*D + blockIdx.z*D + tid];
         VKQ_denominator += KQ_max_scale * meta[l].y;
     }
 
-    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+    dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator;
 }
 
 static void on_no_fattn_vec_case(const int D) {
@@ -671,12 +670,10 @@ static void on_no_fattn_vec_case(const int D) {
     }
 }
 
-// parallel_blocks == 0 is stream-k decomposition
-template <int D, int ncols1, int ncols2, int parallel_blocks, int KQ_stride>
+template <int D, int ncols1, int ncols2, 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,
-    const int warp_size = WARP_SIZE
+    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
+    const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
@@ -748,12 +745,14 @@ void launch_fattn(
         nb23 = nb23*bs*sizeof(half)/ts;
     }
 
+    int parallel_blocks = 1;
+
     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) {
+    if (stream_k) {
         // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup.
         const int max_blocks = 2*nsm;
         const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks;
@@ -769,9 +768,43 @@ void launch_fattn(
 
         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];
-        blocks_num.z = Q->ne[3];
+        GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
+        const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+
+        int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
+        CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
+
+        // parallel_blocks should be at least large enough to achieve max. occupancy for a single wave:
+        parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1);
+
+        // parallel_blocks must not be larger than what the tensor size allows:
+        parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
+
+        // If ntiles_total % blocks_per_wave != 0 then some efficiency is lost due to tail effects.
+        // Test whether parallel_blocks can be set to a higher value for better efficiency.
+        const int blocks_per_wave = nsm * max_blocks_per_sm;
+        int nwaves_best = 0;
+        int efficiency_percent_best = 0;
+        for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KQ; ++parallel_blocks_test) {
+            const int nblocks_total = ntiles_total * parallel_blocks_test;
+            const int nwaves = (nblocks_total + blocks_per_wave - 1) / blocks_per_wave;
+            const int efficiency_percent = 100 * nblocks_total / (nwaves*blocks_per_wave);
+
+            // Stop trying configurations with more waves if we already have good efficiency to avoid excessive overhead.
+            if (efficiency_percent_best >= 90 && nwaves > nwaves_best) {
+                break;
+            }
+
+            if (efficiency_percent > efficiency_percent_best) {
+                nwaves_best = nwaves;
+                efficiency_percent_best = efficiency_percent;
+                parallel_blocks = parallel_blocks_test;
+            }
+        }
+
+        blocks_num.x = ntiles_x;
+        blocks_num.y = parallel_blocks;
+        blocks_num.z = Q->ne[2]*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
@@ -803,7 +836,7 @@ void launch_fattn(
         K_data,
         V_data,
         mask ? ((const char *) mask->data) : nullptr,
-        (parallel_blocks) > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
+        !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
         Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3],
@@ -815,7 +848,7 @@ void launch_fattn(
     );
     CUDA_CHECK(cudaGetLastError());
 
-    if constexpr (parallel_blocks == 0) {
+    if (stream_k) {
         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.x, ncols1, ncols2};
@@ -824,13 +857,14 @@ void launch_fattn(
                 <<<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]);
         }
-    } else if constexpr (parallel_blocks > 1) {
+    } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(D, 1, 1);
-        const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+        const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z);
+        const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);
 
-        flash_attn_combine_results<D, parallel_blocks>
-            <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-            (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+        flash_attn_combine_results<D>
+            <<<blocks_num_combine, block_dim_combine, nbytes_shared_combine, main_stream>>>
+            (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks);
     }
     CUDA_CHECK(cudaGetLastError());
 }

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -970,7 +970,8 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
         fattn_kernel = flash_attn_ext_f16<D, ncols1, ncols2, nwarps, KQ_per_iter, ntiles, use_logit_softcap>;
     }
 
-    launch_fattn<D, ncols1, ncols2, 0, KQ_per_iter>(ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, true, true);
+    launch_fattn<D, ncols1, ncols2, KQ_per_iter>
+        (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, FATTN_KQ_STRIDE, true, true, true);
 }
 
 

ggml/src/ggml-cuda/fattn-tile-f16.cu

@@ -4,7 +4,7 @@
 
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
-template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -58,18 +58,17 @@ static __global__ void flash_attn_tile_ext_f16(
 
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
     const half   * maskh = (const half   *)  mask + ne11*ic0;
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -105,8 +104,7 @@ static __global__ void flash_attn_tile_ext_f16(
 
     __syncthreads();
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         half kqmax_new[ncols/nwarps];
@@ -271,16 +269,16 @@ static __global__ void flash_attn_tile_ext_f16(
             const int i0 = i00 + 2*threadIdx.x;
 
             half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                 dst_val /= __half2half2(kqsum_j);
             }
-            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = __high2float(dst_val);
         }
 
-        if (parallel_blocks != 1 && threadIdx.x == 0) {
-            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else
@@ -288,7 +286,7 @@ static __global__ void flash_attn_tile_ext_f16(
 #endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
 }
 
-template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+template <int cols_per_block, bool use_logit_softcap>
 void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
     switch (Q->ne[0]) {
@@ -296,15 +294,17 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    D             = 64;
             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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } 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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
         } break;
         default: {
             GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -324,37 +324,22 @@ void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_ten
 
     if (Q->ne[1] <= 16) {
         constexpr int cols_per_block = 16;
-        constexpr int parallel_blocks = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 32) {
-        constexpr int cols_per_block = 32;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     constexpr int cols_per_block = 32;
-    constexpr int parallel_blocks = 1;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f16_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     }
 }

ggml/src/ggml-cuda/fattn-tile-f32.cu

@@ -4,7 +4,7 @@
 
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
-template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -58,18 +58,17 @@ static __global__ void flash_attn_tile_ext_f32(
 
     // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
     const half   * maskh = (const half   *)  mask + ne11*ic0;
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
 
@@ -103,8 +102,7 @@ static __global__ void flash_attn_tile_ext_f32(
 
     __syncthreads();
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         float kqmax_new[ncols/nwarps];
@@ -269,17 +267,17 @@ static __global__ void flash_attn_tile_ext_f32(
             const int i0 = i00 + 2*threadIdx.x;
 
             float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                 dst_val.x /= kqsum_j;
                 dst_val.y /= kqsum_j;
             }
-            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
-            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
+            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = dst_val.x;
+            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = dst_val.y;
         }
 
-        if (parallel_blocks != 1 && threadIdx.x == 0) {
-            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        if (gridDim.y != 1 && threadIdx.x == 0) {
+            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else
@@ -287,7 +285,7 @@ static __global__ void flash_attn_tile_ext_f32(
 #endif // FLASH_ATTN_AVAILABLE
 }
 
-template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+template <int cols_per_block, bool use_logit_softcap>
 void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * Q = dst->src[0];
     switch (Q->ne[0]) {
@@ -295,15 +293,17 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
             constexpr int    D             = 64;
             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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } 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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, use_logit_softcap>;
+            launch_fattn<D, cols_per_block, 1, -1>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false);
         } break;
         default: {
             GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
@@ -320,37 +320,22 @@ void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_ten
 
     if (Q->ne[1] <= 16) {
         constexpr int cols_per_block = 16;
-        constexpr int parallel_blocks = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 32) {
-        constexpr int cols_per_block = 32;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+            launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     constexpr int cols_per_block = 32;
-    constexpr int parallel_blocks = 1;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        launch_fattn_tile_f32_64_128<cols_per_block, use_logit_softcap>(ctx, dst);
     }
 }

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -55,17 +55,16 @@ static __global__ void flash_attn_vec_ext_f16(
     constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
     constexpr dequantize_1_f16_t dequantize_1_v = get_dequantize_1_f16(type_V);
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.y              + nb01*ic0;
-    K += nb12*(blockIdx.y / gqa_ratio);
-    V += nb22*(blockIdx.y / gqa_ratio);
+    Q += nb02* blockIdx.z              + nb01*ic0;
+    K += nb12*(blockIdx.z / gqa_ratio);
+    V += nb22*(blockIdx.z / gqa_ratio);
 
     const half * maskh = (const half   *)  mask + ne11*ic0;
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -172,8 +171,7 @@ static __global__ void flash_attn_vec_ext_f16(
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
@@ -283,29 +281,29 @@ static __global__ void flash_attn_vec_ext_f16(
         kqsum[j_VKQ] = warp_reduce_sum((float)kqsum[j_VKQ]);
 
         half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
-        if (parallel_blocks == 1) {
+        if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
     }
 
-    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]);
+    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
    NO_DEVICE_CODE;
 #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>
+template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, type_K, type_V, use_logit_softcap>;
     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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>
@@ -325,65 +323,48 @@ void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
     if (Q->ne[1] == 1) {
-        constexpr int cols_per_block  = 1;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 1;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] == 2) {
-        constexpr int cols_per_block  = 2;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 2;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block  = 4;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block  = 8;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block  = 8;
-    constexpr int parallel_blocks = 1;
+    constexpr int cols_per_block = 8;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     }
 }
 

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "fattn-common.cuh"
 
-template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
@@ -55,16 +55,15 @@ static __global__ void flash_attn_vec_ext_f32(
     constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
     constexpr dequantize_1_f32_t dequantize_1_v = get_dequantize_1_f32(type_V);
 
-    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
-    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.y              + nb01*ic0;
-    K += nb12*(blockIdx.y / gqa_ratio);
-    V += nb22*(blockIdx.y / gqa_ratio); // K and V have same shape
+    Q += nb02* blockIdx.z              + nb01*ic0;
+    K += nb12*(blockIdx.z / gqa_ratio);
+    V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
     const half * maskh = (const half   *)  mask + ne11*ic0;
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
     constexpr int nwarps = D / WARP_SIZE;
@@ -167,8 +166,7 @@ static __global__ void flash_attn_vec_ext_f32(
 
     float VKQ[ncols] = {0.0f};
 
-    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) {
         // Calculate KQ tile and keep track of new maximum KQ values:
 
         float kqmax_new_arr[ncols];
@@ -268,29 +266,29 @@ static __global__ void flash_attn_vec_ext_f32(
         kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
 
         float dst_val = VKQ[j_VKQ];
-        if (parallel_blocks == 1) {
+        if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
-        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
+        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
     }
 
-    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]);
+    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = 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>
+template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
 void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, type_K, type_V, use_logit_softcap>;
     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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
 }
 
 template <int D, ggml_type type_K, ggml_type type_V>
@@ -307,65 +305,48 @@ void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
     if (Q->ne[1] == 1) {
-        constexpr int cols_per_block  = 1;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 1;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] == 2) {
-        constexpr int cols_per_block  = 2;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 2;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
     if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block  = 4;
-        constexpr int parallel_blocks = 4;
+        constexpr int cols_per_block = 4;
         if (logit_softcap == 0.0f) {
             constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         } else {
             constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
         }
         return;
     }
 
-    if (Q->ne[1] <= 8) {
-        constexpr int cols_per_block  = 8;
-        constexpr int parallel_blocks = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block  = 8;
-    constexpr int parallel_blocks = 1;
+    constexpr int cols_per_block = 8;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     } else {
         constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
     }
 }
 

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -18,7 +18,7 @@ namespace wmma = rocwmma;
 #endif // FP16_MMA_AVAILABLE
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
-template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
+template<int D, int ncols, int nwarps, int VKQ_stride, typename KQ_acc_t, bool use_logit_softcap>
 __launch_bounds__(nwarps*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
@@ -67,8 +67,7 @@ static __global__ void flash_attn_ext_f16(
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-    const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.
-    const int ip  =        blockIdx.x % parallel_blocks;  // Index in group of blocks running for the same column in parallel.
+    const int ic0 = ncols*blockIdx.x; // Index of the first Q/QKV column to work on.
 
     static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE.");
     static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16.");
@@ -91,16 +90,16 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
 
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f   = (const float *) (Q + nb02* blockIdx.y              + nb01*ic0);
-    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.y / gqa_ratio));
-    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const float * Q_f   = (const float *) (Q + nb02* blockIdx.z              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.z / gqa_ratio));
+    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
     const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
     const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
 
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
     const half2 slope2 = make_half2(slopef, slopef);
 
@@ -176,7 +175,7 @@ static __global__ void flash_attn_ext_f16(
     __syncthreads();
 
     // Iterate over ne11 == previous tokens:
-    for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) {
+    for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) {
         // Calculate tile of KQ:
 #pragma unroll
         for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
@@ -395,7 +394,7 @@ static __global__ void flash_attn_ext_f16(
         if (ic0 + j_VKQ >= ne01) {
             return;
         }
-        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
 
         float KQ_rowsum_j;
         if (std::is_same<KQ_acc_t, float>::value) {
@@ -411,13 +410,13 @@ static __global__ void flash_attn_ext_f16(
                 break;
             }
             float dst_val = VKQ[j_VKQ*D_padded + i];
-            if (parallel_blocks == 1) {
+            if (gridDim.y == 1) {
                 dst_val /= KQ_rowsum_j;
             }
-            dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val;
+            dst[j_dst*gridDim.z*D + blockIdx.z*D + i] = dst_val;
         }
 
-        if (parallel_blocks == 1 || threadIdx.x != 0) {
+        if (gridDim.y == 1 || threadIdx.x != 0) {
             continue;
         }
 
@@ -428,7 +427,7 @@ static __global__ void flash_attn_ext_f16(
             dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
         }
         dst_meta_val.y = KQ_rowsum_j;
-        dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val;
+        dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = dst_meta_val;
     }
 #else
    NO_DEVICE_CODE;
@@ -462,60 +461,26 @@ static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 template <int D, int cols_per_block, typename KQ_acc_t>
 void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q   = dst->src[0];
 
     constexpr int nwarps = 4;
 
     constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
-    const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
-    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
     const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
 
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
-    if (4*blocks_num_pb1 < 2*nsm) {
-        constexpr int parallel_blocks = 4;
-        fattn_kernel_t fattn_kernel;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            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>;
-        } else {
-            constexpr bool use_logit_softcap = true;
-            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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
-        return;
-    }
-    if (2*blocks_num_pb1 < 2*nsm) {
-        constexpr int parallel_blocks = 2;
-        fattn_kernel_t fattn_kernel;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            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>;
-        } else {
-            constexpr bool use_logit_softcap = true;
-            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, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
-        return;
-    }
-    constexpr int parallel_blocks = 1;
     fattn_kernel_t fattn_kernel;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
         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>;
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
     } else {
         constexpr bool use_logit_softcap = true;
         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>;
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
     }
-    launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, 0, true, true, warp_size);
+    launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
 }
 
 void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/fattn.cu

@@ -281,13 +281,13 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     if (!fp16_mma_available(cc)) {
         if (prec == GGML_PREC_DEFAULT) {
-            if (Q->ne[1] <= 8) {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
                 ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
             } else {
                 ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
             }
         } else {
-            if (Q->ne[1] <= 8) {
+            if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
                 ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
             } else {
                 ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
@@ -296,17 +296,17 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
         return;
     }
 
-    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) {
+    const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
+    const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
+    const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion;
+    const bool can_use_vector_kernel = (Q->ne[0] % (2*warp_size) == 0) && (prec == GGML_PREC_DEFAULT || Q->ne[0] <= 128);
+    if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
         if (prec == GGML_PREC_DEFAULT) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-            return;
-        } else if(Q->ne[0] <= 128) {
+        } else {
             ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-            return;
         }
+        return;
     }
 
     // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -3230,6 +3230,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 #ifndef FLASH_ATTN_AVAILABLE
             return false;
 #endif // FLASH_ATTN_AVAILABLE
+            if (op->src[0]->ne[3] != 1) {
+                return false;
+            }
             if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                 return false;
             }

ggml/src/ggml-cuda/vendors/hip.h

@@ -129,6 +129,7 @@
 #define cudaGraph_t hipGraph_t
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
 #define __trap() do { abort(); __builtin_unreachable(); } while(0)
 #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
 #define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED

ggml/src/ggml-cuda/vendors/musa.h

@@ -134,5 +134,6 @@
 #define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
 #define cudaStreamBeginCapture musaStreamBeginCapture
 #define cudaStreamEndCapture musaStreamEndCapture
+#define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor
 
 typedef mt_bfloat16 nv_bfloat16;

ggml/src/ggml-sycl/CMakeLists.txt

@@ -23,6 +23,38 @@ ggml_add_backend_library(ggml-sycl
                          ../../include/ggml-sycl.h
                         )
 
+find_package(DNNL)
+set(GGML_SYCL_DNNL 0)
+if(DNNL_FOUND)
+    if (DEFINED ENV{ONEAPI_ROOT} AND NOT DEFINED DNNL_GPU_VENDOR)
+        # Assuming oneDNN packaged with oneapi release is used which
+        # supports only intel target
+        set(DNNL_GPU_VENDOR "INTEL")
+        if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL")
+            message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target")
+        endif()
+    endif()
+
+    # Verify oneDNN was compiled for the same target as llama
+    if("${GGML_SYCL_TARGET}" STREQUAL "${DNNL_GPU_VENDOR}")
+        target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
+        set(GGML_SYCL_DNNL 1)
+        get_target_property(CONFIGS DNNL::dnnl IMPORTED_CONFIGURATIONS)
+        foreach(CONFIG ${CONFIGS})
+            get_target_property(DNNL_LIB DNNL::dnnl IMPORTED_LOCATION_${CONFIG})
+            message(STATUS "Found oneDNN: ${DNNL_LIB}")
+        endforeach()
+    else()
+        message(WARNING
+            "oneDNN must be compiled for the same target as llama.cpp.
+             llama.cpp: ${GGML_SYCL_TARGET}, oneDNN: ${DNNL_GPU_VENDOR}.
+             Disabling oneDNN support.")
+    endif()
+else()
+    message(STATUS "oneDNN not found, disabling oneDNN support")
+endif()
+target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_DNNL=${GGML_SYCL_DNNL})
+
 if (GGML_SYCL_F16)
     if (GGML_SYCL_TARGET STREQUAL "AMD")
         message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
@@ -48,18 +80,6 @@ file(GLOB   GGML_HEADERS_SYCL "*.hpp")
 file(GLOB   GGML_SOURCES_SYCL "*.cpp")
 target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
 
-find_package(DNNL)
-message("-- DNNL found:" ${DNNL_FOUND})
-
-if (GGML_SYCL_TARGET STREQUAL "INTEL")
-    add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
-else()
-    add_compile_definitions(GGML_SYCL_DNNL=0)
-endif()
-
-if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
-    target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
-endif()
 
 if (WIN32)
     find_package(IntelSYCL REQUIRED)

ggml/src/ggml-sycl/common.hpp

@@ -170,7 +170,6 @@ static size_t g_scratch_offset = 0;
 int get_current_device_id();
 
 inline dpct::err0 ggml_sycl_set_device(const int device) try {
-
   int current_device_id;
   SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id()));
 
@@ -242,6 +241,14 @@ struct ggml_sycl_pool_alloc {
         }
     }
 
+    T * realloc(size_t size) {
+        GGML_ASSERT(pool != nullptr);
+        if (ptr)
+            pool->free(ptr, actual_size);
+        ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
+        return ptr;
+    }
+
     // size is in number of elements
     T * alloc(size_t size) {
         GGML_ASSERT(pool != nullptr);
@@ -371,10 +378,29 @@ struct ggml_backend_sycl_context {
     dnnl::stream stream_dnnl() {
         return stream_dnnl(device, 0);
     }
+    dnnl::memory get_scratchpad_mem(const dnnl::memory::desc & scratchpad_md,
+                                    const dnnl::engine & eng, const queue_ptr q) {
+        ggml_sycl_pool_alloc<uint8_t> * pool;
+        auto it = scratchpad_map.find(q);
+        if (it == scratchpad_map.end()) {
+            scratchpad_map[q] = std::make_unique<ggml_sycl_pool_alloc<uint8_t>>(this->pool());
+            pool = scratchpad_map[q].get();
+        } else {
+            pool = it->second.get();
+        }
+
+        size_t scratchpad_size = scratchpad_md.get_size();
+        if (scratchpad_size > pool->actual_size) {
+            pool->realloc(scratchpad_size);
+        }
+        void * mem_ptr = pool->get();
+        return dnnl::memory(scratchpad_md, eng, mem_ptr);
+    }
 #endif
 
     // pool
     std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES];
+    std::unordered_map<sycl::queue *, std::unique_ptr<ggml_sycl_pool_alloc<uint8_t>>> scratchpad_map;
 
     std::unique_ptr<ggml_sycl_pool> host_pools[GGML_SYCL_MAX_DEVICES];
 

ggml/src/ggml-sycl/gemm.hpp

@@ -13,9 +13,6 @@
 #ifndef GGML_SYCL_GEMM_HPP
 #define GGML_SYCL_GEMM_HPP
 
-#include <fstream>
-#include <iostream>
-
 #include "ggml-sycl.h"
 
 #if GGML_SYCL_DNNL
@@ -35,62 +32,34 @@ public:
         else static_assert(0);
     }
 
-    static inline void row_gemm(sycl::queue& q, bool a_trans,
-        bool b_trans, int m, int n, int k,
-        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
-    {
-        // Get the device associated with the queue
-        sycl::device dev = q.get_device();
-        // Get the context associated with the queue
-        sycl::context ctx = q.get_context();
-        const dnnl::engine eng = dnnl::sycl_interop::make_engine(dev, ctx);
-        const dnnl::stream stream = dnnl::sycl_interop::make_stream(eng, q);
+    static inline void row_gemm(ggml_backend_sycl_context & ctx, bool a_trans, bool b_trans, int m, int n, int k,
+                                const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
+        auto stream = ctx.stream_dnnl(q);
+        auto eng = ctx.engine_dnnl(q);
         dnnl::memory::dims a_dims = { m, k };
         dnnl::memory::dims b_dims = { k, n };
         dnnl::memory::dims c_dims = { m, n };
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
         const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
+        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::ab);
+
+        dnnl::primitive_attr primitive_attr;
+        primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+
         auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
         auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
-        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
+        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md, primitive_attr);
         auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
 
-        // Create the primitive.
+        auto scratchpad_md = matmul_pd.scratchpad_desc();
+        auto scratchpad_mem = ctx.get_scratchpad_mem(scratchpad_md, eng, q);
         auto matmul_prim = dnnl::matmul(matmul_pd);
-        // Primitive arguments.
-        std::unordered_map<int, dnnl::memory> matmul_args;
-        matmul_args.insert({ DNNL_ARG_SRC, a_mem });
-        matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
-        matmul_args.insert({ DNNL_ARG_DST, c_mem });
-
-        matmul_prim.execute(stream, matmul_args);
-    }
-
-
-    static inline void row_gemm(const dnnl::stream& stream, bool a_trans,
-        bool b_trans, int m, int n, int k,
-        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
-    {
-        auto const eng = stream.get_engine();
-        dnnl::memory::dims a_dims = { m, k };
-        dnnl::memory::dims b_dims = { k, n };
-        dnnl::memory::dims c_dims = { m, n };
-        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
-        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
-        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
-        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
-        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
-        auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
-
-        // Create the primitive.
-        auto matmul_prim = dnnl::matmul(matmul_pd);
-        // Primitive arguments.
+
         std::unordered_map<int, dnnl::memory> matmul_args;
         matmul_args.insert({ DNNL_ARG_SRC, a_mem });
         matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
         matmul_args.insert({ DNNL_ARG_DST, c_mem });
+        matmul_args.insert({ DNNL_ARG_SCRATCHPAD, scratchpad_mem });
 
         matmul_prim.execute(stream, matmul_args);
     }

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -2058,9 +2058,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
         to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
 #else
-        auto dnnl_stream = ctx.stream_dnnl(stream);
-        DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
-            src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>());
+        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ptr,
+                                  DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
+                                  dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
         const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
         to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
 #endif
@@ -2099,9 +2099,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
             dst_dd_i, ldc)));
 #    endif
 #else
-        auto dnnl_stream = ctx.stream_dnnl(stream);
-         DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
-            src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
+        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i,
+                                  DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
+                                  dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
 #endif
     }
     GGML_UNUSED(dst);

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -8436,8 +8436,12 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
 
+    uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false);
+        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+            total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        }
     }
     if (ctx->device->need_compiles) {
         ggml_vk_load_shaders(ctx->device);
@@ -8458,17 +8462,27 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     bool first_node_in_batch = true; // true if next node will be first node in a batch
     int submit_node_idx = 0; // index to first node in a batch
 
-    // Submit work every nodes_per_submit nodes to overlap CPU cmdbuffer generation with GPU execution.
-    // Start with a smaller count to get work submitted right away, and increase it after each submit.
-    int nodes_per_submit = 20;
+    // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
+    // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
+    // (and scaled down based on model size, so smaller models submit earlier).
+    // Also submit at least every 100 nodes, in case there are workloads without as much matmul.
+    int nodes_per_submit = 100;
     int submitted_nodes = 0;
     int submit_count = 0;
+    uint64_t mul_mat_bytes = 0;
+    uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), total_mat_mul_bytes / 40u);
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (first_node_in_batch) {
             submit_node_idx = i;
         }
 
-        bool submit = (submitted_nodes >= nodes_per_submit) || (i == last_node);
+        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+            mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        }
+
+        bool submit = (submitted_nodes >= nodes_per_submit) ||
+                      (mul_mat_bytes >= mul_mat_bytes_per_submit) ||
+                      (i == last_node);
 
         bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit);
 
@@ -8485,13 +8499,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         if (submit) {
             first_node_in_batch = true;
             submitted_nodes = 0;
-            switch (submit_count) {
-            case 0:
-                nodes_per_submit = 50;
-                break;
-            default:
-                nodes_per_submit = 100;
-                break;
+            mul_mat_bytes = 0;
+            if (submit_count < 3) {
+                mul_mat_bytes_per_submit *= 2;
             }
             submit_count++;
         }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp

@@ -82,8 +82,8 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1]));
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
-    const i8vec2 v0 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2]);
-    const i8vec2 v1 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2 + 1]);
+    const i8vec2 v0 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2])).xy; // vec4 used due to #12147
+    const i8vec2 v1 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2 + 1])).xy;
     return vec4(v0.x, v0.y, v1.x, v1.y);
 }
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -311,8 +311,8 @@ float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords
     const float16_t d = bl.block.d;
     const uint idx = coordInBlock[1];
 
-    const uint ib32 = idx / 32;
-    const uint ib8 = idx / 8;
+    const uint ib32 = (idx & 0xE0) >> 5;
+    const uint ib8 = (idx & 0xF8) >> 3;
 
     const uint qh = bl.block.qh[ib32];
     const uint qs = bl.block.qs[ib8];
@@ -330,14 +330,20 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1
    block_iq1_m block;
 };
 
+layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufIQ1_M_packed64 {
+   block_iq1_m_packed64 block;
+};
+
 float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
-    const u16vec4 scales = u16vec4(bl.block.scales[0], bl.block.scales[1], bl.block.scales[2], bl.block.scales[3]) >> 12;
-    const float16_t d = uint16BitsToHalf(scales.x | (scales.y << 4) | (scales.z << 8) | (scales.w << 12));
+    decodeBufIQ1_M_packed64 bl64 = decodeBufIQ1_M_packed64(bl);
     const uint idx = coordInBlock[1];
 
-    const uint ib8 = idx / 8;
-    const uint ib16 = idx / 16;
+    uvec2 scales = unpack32(bl64.block.scales);
+    const float16_t d = uint16BitsToHalf(uint16_t(((scales.x & 0xF000) >> 12) | ((scales.x & 0xF0000000) >> 24) | ((scales.y & 0xF000) >> 4) | ((scales.y & 0xF0000000) >> 16)));
+
+    const uint ib8 = (idx & 0xF8) >> 3;
+    const uint ib16 = (idx & 0xF0) >> 4;
     const int i8 = int(idx % 8);
     const uint sc = bl.block.scales[ib8 / 8];
     const uint qs = bl.block.qs[ib8];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp

@@ -19,8 +19,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
         const float db = d * (0.5 + scale) * 0.25;
 
         const uint qh = data_a[ibi].qh[ib32];
-        const u8vec2 qs16 = unpack8(data_a_packed16[ibi].qs[itid]);
-        const u8vec2 sign16 = unpack8(data_a_packed16[ibi].qs[QUANT_K / 16 + itid]);
+        const u8vec2 qs16 = unpack8(uint32_t(data_a_packed16[ibi].qs[itid])).xy; // vec4 used due to #12147
+        const u8vec2 sign16 = unpack8(uint32_t(data_a_packed16[ibi].qs[QUANT_K / 16 + itid])).xy;
         [[unroll]] for (uint l = 0; l < 2; ++l) {
             const uint8_t sign = sign16[l];
             const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300);

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp

@@ -21,7 +21,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
             sum[j] = 0.0;
         }
         [[unroll]] for (uint l = 0; l < 4; ++l) {
-            const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]);
+            const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147
             const uint sign = data_a[ibi].signs[4 * ib32 + l];
             const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
             const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)]));

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -336,8 +336,8 @@ void main() {
             const uint iqs = idx & 0x07;
 
             const float d = float(data_a_packed16[ib].d);
-            const i8vec2 v0 = unpack8(data_a_packed16[ib].qs[2*iqs]);
-            const i8vec2 v1 = unpack8(data_a_packed16[ib].qs[2*iqs + 1]);
+            const i8vec2 v0 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs])).xy; // vec4 used due to #12147
+            const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy;
             const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d;
 
             buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
@@ -544,7 +544,7 @@ void main() {
             const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
             const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
             const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
 
             buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -564,7 +564,7 @@ void main() {
             const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
             const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
             const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
 
             buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -586,7 +586,7 @@ void main() {
             const float db = d * 0.25 * (0.5 + scale);
             const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
             const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(uint32_t(grid)).xy); // vec4 used due to #12147
 
             buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -611,7 +611,7 @@ void main() {
             const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
             const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
             const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
 
             buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -631,7 +631,7 @@ void main() {
             const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
             const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
             const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
-            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
 
             buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);

ggml/src/ggml-vulkan/vulkan-shaders/types.comp

@@ -2,6 +2,7 @@
 #if !defined(GGML_TYPES_COMP)
 #define GGML_TYPES_COMP
 
+#extension GL_EXT_shader_explicit_arithmetic_types_int64 : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
@@ -312,6 +313,12 @@ struct block_iq1_m {
     uint16_t scales[QUANT_K_IQ1_M/64];
 };
 
+struct block_iq1_m_packed64 {
+    uint64_t  qs[QUANT_K_IQ1_M/8/8];
+    uint64_t  qh[QUANT_K_IQ1_M/16/8];
+    uint64_t scales;
+};
+
 #if defined(DATA_A_IQ1_S)
 #define QUANT_K QUANT_K_IQ1_S
 #define QUANT_R QUANT_R_IQ1_S

gguf-py/gguf/vocab.py

@@ -154,7 +154,12 @@ class SpecialVocab:
             return True
         with open(tokenizer_config_file, encoding = 'utf-8') as f:
             tokenizer_config = json.load(f)
-        chat_template = tokenizer_config.get('chat_template')
+        chat_template_alt = None
+        chat_template_file = path / 'chat_template.json'
+        if chat_template_file.is_file():
+            with open(chat_template_file, encoding = 'utf-8') as f:
+                chat_template_alt = json.load(f).get('chat_template')
+        chat_template = tokenizer_config.get('chat_template', chat_template_alt)
         if chat_template is None or isinstance(chat_template, (str, list)):
             self.chat_template = chat_template
         else:

src/llama-context.cpp

@@ -1153,6 +1153,7 @@ int llama_context::encode(llama_batch & inp_batch) {
         // remember the sequence ids used during the encoding - needed for cross attention later
         cross.seq_ids_enc.resize(n_tokens);
         for (int32_t i = 0; i < n_tokens; i++) {
+            cross.seq_ids_enc[i].clear();
             for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
                 llama_seq_id seq_id = ubatch.seq_id[i][s];
                 cross.seq_ids_enc[i].insert(seq_id);

src/llama-model.cpp

@@ -271,19 +271,32 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
-    // add extra buffer types
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
-    if (ggml_backend_dev_get_extra_bufts_fn) {
-        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
-        while (extra_bufts && *extra_bufts) {
-            buft_list.emplace_back(cpu_dev, *extra_bufts);
-            ++extra_bufts;
+    bool has_gpu_device = false;
+    for (auto * dev : devices) {
+        if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            has_gpu_device = true;
+            break;
         }
     }
 
+    // add extra buffer types, only if no GPU device is present
+    // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
+    if (!has_gpu_device) {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_list.emplace_back(cpu_dev, *extra_bufts);
+                ++extra_bufts;
+            }
+        }
+    } else {
+        LLAMA_LOG_WARN("%s: disabling extra buffer types (i.e. repacking) since a GPU device is available\n", __func__);
+    }
+
     // add a host buffer type
     // storing the tensors in a host buffer is useful when the processing of large batches
     // is offloaded to a GPU device, since it reduces the time spent on data transfers
@@ -2020,7 +2033,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     // output
                     output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
                     output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
-                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -2205,9 +2223,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
                         // optional bias tensors
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, 0);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -2324,7 +2342,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
                         layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), { n_embd }, 0);
 
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, TENSOR_NOT_REQUIRED);
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
                             layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias",   i), {n_embd}, 0);
@@ -2381,7 +2399,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     // output
                     output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
                     output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
-                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -2407,7 +2430,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_CODESHELL:
                 {
-                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if tok embd is NULL, init from output
+                    if (tok_embd == NULL) {
+                        tok_embd = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     // output
                     output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -3200,16 +3228,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
                             layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
                             layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
-                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
                         }
 
                         layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
@@ -3320,12 +3348,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {time_mix_extra_dim, n_embd, 5}, 0);
 
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
-                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, TENSOR_NOT_REQUIRED);
                         GGML_ASSERT(!(layer.time_mix_lerp_fused == NULL && layer.time_mix_lerp_w == NULL));
 
                         layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, 0);
@@ -3355,7 +3383,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, TENSOR_NOT_REQUIRED);
                     output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
 
                     const int time_mix_extra_dim = hparams.time_mix_extra_dim;
@@ -3381,7 +3409,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
                         layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0);
 
-                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_decay = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY, "weight", i), {n_embd}, 0);
                         layer.time_mix_decay_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W1, "weight", i), {n_embd, time_decay_extra_dim}, 0);
                         layer.time_mix_decay_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W2, "weight", i), {time_decay_extra_dim, attn_hidden_size}, 0);
@@ -3390,9 +3418,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_gate = create_tensor(tn(LLM_TENSOR_TIME_MIX_GATE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         // optional bias tensors
-                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, TENSOR_NOT_REQUIRED);
 
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
@@ -3513,8 +3541,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0);
                         }
 
-                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED);
 
                         try {
                             layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0);
@@ -3531,8 +3559,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
 
-                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@@ -6178,16 +6206,25 @@ struct llm_build_qwen2moe : public llm_graph_context {
             {
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
 
                 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
 
                 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);

tests/test-backend-ops.cpp

@@ -259,6 +259,10 @@ static std::string var_to_str(ggml_type type) {
     return ggml_type_name(type);
 }
 
+static std::string var_to_str(ggml_prec prec) {
+    return prec == GGML_PREC_F32 ? "f32" : "def";
+}
+
 static std::string var_to_str(ggml_op_pool pool) {
     switch (pool) {
         case GGML_OP_POOL_AVG:  return "avg";
@@ -3206,11 +3210,12 @@ struct test_flash_attn_ext : public test_case {
     const float max_bias; // ALiBi
     const float logit_softcap; // Gemma 2
 
+    const ggml_prec prec;
     const ggml_type type_KV;
     std::array<int32_t, 4> permute;
 
     std::string vars() override {
-        return VARS_TO_STR10(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV, permute);
+        return VARS_TO_STR11(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
     }
 
     double max_nmse_err() override {
@@ -3225,9 +3230,9 @@ struct test_flash_attn_ext : public test_case {
     }
 
     test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
-                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16,
-                        std::array<int32_t, 4> permute = {0, 1, 2, 3})
-        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV), permute(permute) {}
+                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
+                        ggml_type type_KV = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3})
+        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
@@ -3261,6 +3266,7 @@ struct test_flash_attn_ext : public test_case {
         }
 
         ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
+        ggml_flash_attn_ext_set_prec(out, prec);
         ggml_set_name(out, "out");
 
         return out;
@@ -4376,11 +4382,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                             for (int kv : { 512, 1024, }) {
                                 if (nr != 1 && kv != 512) continue;
                                 for (int nb : { 1, 3, 32, 35, }) {
-                                    for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
-                                        test_cases.emplace_back(new test_flash_attn_ext(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV));
-                                        // run fewer test cases permuted
-                                        if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
-                                            test_cases.emplace_back(new test_flash_attn_ext(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV, {0, 2, 1, 3}));
+                                    for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
+                                        if (hs != 128 && prec == GGML_PREC_DEFAULT) continue;
+                                        for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
+                                            test_cases.emplace_back(new test_flash_attn_ext(
+                                                hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
+                                            // run fewer test cases permuted
+                                            if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+                                                test_cases.emplace_back(new test_flash_attn_ext(
+                                                    hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                            }
                                         }
                                     }
                                 }