fix(vulkan): refresh ggml for AMD submission handling (#3482)

Refresh the pinned llama.cpp revision with AMD Vulkan submission batching and DeviceLost diagnostics, adapt the sampler API, and document Windows troubleshooting. Refs #3479.

Author: LauraGPT

runtime/llama.cpp/CMakeLists.txt

@@ -34,7 +34,9 @@ set(LLAMA_BUILD_SERVER   OFF CACHE BOOL "" FORCE)
 set(LLAMA_CURL           OFF CACHE BOOL "" FORCE)
 FetchContent_Declare(llama
     GIT_REPOSITORY https://github.com/ggml-org/llama.cpp.git
-    GIT_TAG        8086439a4cea94c71a5dfb8fe4ad1546aebd640f)
+    # Includes the Vulkan submission-batching and DeviceLost diagnostics fixes
+    # needed by recent AMD drivers (ggml-org/llama.cpp#26371).
+    GIT_TAG        803b7fcae893e9caaee3921779628fef83ac0965)
 FetchContent_GetProperties(llama)
 if(NOT llama_POPULATED)
     FetchContent_Populate(llama)

runtime/llama.cpp/README.md

@@ -187,6 +187,28 @@ cmake --build build-vulkan --config Release --target llama-funasr-sensevoice
 `--backend cpu` remains the default. The Windows Vulkan package currently
 accelerates SenseVoiceSmall only, matching the Linux Vulkan package.
 
+#### AMD Windows troubleshooting
+
+Recent AMD drivers can report `VK_ERROR_DEVICE_LOST` or terminate the process
+when a graph submission exceeds the driver's timeout. The pinned ggml revision
+reduces submission sizes on smaller AMD GPUs and fixes the batching threshold.
+To force smaller submissions or collect the last submitted tensors for a bug
+report, run from PowerShell with:
+
+```powershell
+$env:GGML_VK_MAX_NODES_PER_SUBMIT = "16"
+$env:GGML_VK_SERIALIZE_SUBMISSIONS = "1"
+.\llama-funasr-sensevoice.exe `
+  -m sensevoice-small-f16.gguf --vad fsmn-vad.gguf -a sample.wav --backend vulkan
+```
+
+Try `8` or `1` if `16` still triggers a driver timeout. Include the complete
+stderr output, GPU model, and driver version when reporting a failure. Remove
+the variables after diagnosis because serial submissions can reduce throughput.
+Use `--backend cpu` as the reliable fallback on affected driver/device pairs.
+The current SenseVoiceSmall graph does not create a flash-attention operation,
+so a `--no-flash-attn` switch would not change this execution path.
+
 ## Build (shared)
 ```bash
 git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp

runtime/llama.cpp/fun-asr-nano/funasr-cli/funasr-cli.cpp

@@ -208,7 +208,7 @@ int main(int argc,char**argv){
     llama_context*ctx=llama_init_from_model(model,cp);
     if(!ctx){fprintf(stderr,"failed to create llama context\n");llama_model_free(model);return 1;}
     auto sp=llama_sampler_chain_default_params(); llama_sampler*smpl=llama_sampler_chain_init(sp);
-    if(rep!=1.0f) llama_sampler_chain_add(smpl,llama_sampler_init_penalties(256,rep,0.0f,0.0f));
+    if(rep!=1.0f) llama_sampler_chain_add(smpl,llama_sampler_init_penalties(llama_vocab_n_tokens(vocab),256,rep,0.0f,0.0f));
     llama_sampler_chain_add(smpl,llama_sampler_init_greedy());
 
     const char*prefix="<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n语音转写：";