[Bug] Speculative decoding small draft doesn't work on macOS

[vlbosch]

🐛 Bug

I tried to use Mistral Small 7B Instruct v0.3 as draft model for Mistral Large 2407. When not served using "--mode server", the model(s) never respond. I think that's because only CPU is used, instead of GPU as well. When serving with "--mode server", I see that the first token is streamed in the frontend, but then I get the following error: Check failed: (!mstates[i]->draft_output_tokens.empty()) is false.

To Reproduce

Steps to reproduce the behavior:

1. Download Mistral Large 2407
2. Quantize model and gen config
3. Run Mistral Large to see if it works standalone
4. Run the speculative decoding with: python -m mlc_llm serve /Users/USER/LLM/Mistral-Large-Instruct-2407-MLC --additional-models "HF://mlc-ai/Mistral-7B-Instruct-v0.3-q4f16_1-MLC" --speculative-mode small_draft --port 9999 --device metal --mode server
5. First token is streamed, then error message

USER@MBPM3MVLB ~ % python -m mlc_llm serve /Users/USER/LLM/Mistral-Large-Instruct-2407-MLC --additional-models "HF://mlc-ai/Mistral-7B-Instruct-v0.3-q4f16_1-MLC" --speculative-mode small_draft --port 9999 --device metal --mode server
[2024-09-16 08:50:13] INFO auto_device.py:79: Found device: metal:0
[2024-09-16 08:50:13] INFO jit.py:43: MLC_JIT_POLICY = ON. Can be one of: ON, OFF, REDO, READONLY
[2024-09-16 08:50:13] INFO jit.py:158: Using cached model lib: /Users/USER/.cache/mlc_llm/model_lib/3826dfed383847636248c8e5e540102b.dylib
[2024-09-16 08:50:13] INFO download_cache.py:227: Downloading model from HuggingFace: HF://mlc-ai/Mistral-7B-Instruct-v0.3-q4f16_1-MLC
[2024-09-16 08:50:13] INFO download_cache.py:29: MLC_DOWNLOAD_CACHE_POLICY = ON. Can be one of: ON, OFF, REDO, READONLY
[2024-09-16 08:50:13] INFO download_cache.py:166: Weights already downloaded: /Users/USER/.cache/mlc_llm/model_weights/hf/mlc-ai/Mistral-7B-Instruct-v0.3-q4f16_1-MLC
[2024-09-16 08:50:13] INFO jit.py:43: MLC_JIT_POLICY = ON. Can be one of: ON, OFF, REDO, READONLY
[2024-09-16 08:50:13] INFO jit.py:158: Using cached model lib: /Users/USER/.cache/mlc_llm/model_lib/7bbcaf068957bbf173dbd8ad644faea6.dylib
[2024-09-16 08:50:13] INFO engine_base.py:192: The selected engine mode is server. We use as much GPU memory as possible (within the limit of gpu_memory_utilization).
[2024-09-16 08:50:13] INFO engine_base.py:200: If you have low concurrent requests and want to use less GPU memory, please select mode "local".
[2024-09-16 08:50:13] INFO engine_base.py:205: If you don't have concurrent requests and only use the engine interactively, please select mode "interactive".
[08:50:13] /Users/catalyst/Workspace/mlc-ai-package-self-runner/_work/package/package/mlc-llm/cpp/serve/config.cc:688: Under mode "local", max batch size will be set to 4, max KV cache token capacity will be set to 8192, prefill chunk size will be set to 2048.
[08:50:13] /Users/catalyst/Workspace/mlc-ai-package-self-runner/_work/package/package/mlc-llm/cpp/serve/config.cc:688: Under mode "interactive", max batch size will be set to 1, max KV cache token capacity will be set to 32768, prefill chunk size will be set to 2048.
[08:50:13] /Users/catalyst/Workspace/mlc-ai-package-self-runner/_work/package/package/mlc-llm/cpp/serve/config.cc:688: Under mode "server", max batch size will be set to 80, max KV cache token capacity will be set to 32768, prefill chunk size will be set to 2048.
[08:50:13] /Users/catalyst/Workspace/mlc-ai-package-self-runner/_work/package/package/mlc-llm/cpp/serve/config.cc:769: The actual engine mode is "server". So max batch size is 80, max KV cache token capacity is 32768, prefill chunk size is 2048.
[08:50:13] /Users/catalyst/Workspace/mlc-ai-package-self-runner/_work/package/package/mlc-llm/cpp/serve/config.cc:774: Estimated total single GPU memory usage: 86697.674 MB (Parameters: 69664.656 MB. KVCache: 15602.123 MB. Temporary buffer: 1430.894 MB). The actual usage might be slightly larger than the estimated number.
[08:50:13] /Users/catalyst/Workspace/mlc-ai-package-self-runner/_work/package/package/mlc-llm/cpp/serve/engine.cc:365: Warning: Hybrid prefill mode fallbacks to chunked prefill, due to speculative mode is enabled and not implemented with hybrid prefill yet.
INFO: Started server process [69315]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:9999 (Press CTRL+C to quit)
INFO: 127.0.0.1:58406 - "POST /v1/chat/completions HTTP/1.1" 200 OK
libc++abi: terminating due to uncaught exception of type tvm::runtime::InternalError: [08:50:41] /Users/catalyst/Workspace/mlc-ai-package-self-runner/_work/package/package/mlc-llm/cpp/serve/engine_actions/batch_draft.cc:151: InternalError: Check failed: (!mstates[i]->draft_output_tokens.empty()) is false:
Stack trace:

zsh: abort python -m mlc_llm serve /Users/USER/LLM/Mistral-Large-Instruct-2407-MLC

Expected behavior

The model streams the output to the provided prompt.

Environment

• Platform (e.g. WebGPU/Vulkan/IOS/Android/CUDA): Macbook Pro
• Operating system (e.g. Ubuntu/Windows/MacOS/...): macOS Sequoia
• Device (e.g. iPhone 12 Pro, PC+RTX 3090, ...): M3 Max
• How you installed MLC-LLM (conda, source): conda with pip install
• How you installed TVM-Unity (pip, source): pip
• Python version (e.g. 3.10): 3.12
• GPU driver version (if applicable): -
• CUDA/cuDNN version (if applicable): -
• TVM Unity Hash Tag (python -c "import tvm; print('\n'.join(f'{k}: {v}' for k, v in tvm.support.libinfo().items()))", applicable if you compile models):
  "USE_NVTX: OFF
  USE_GTEST: AUTO
  SUMMARIZE: OFF
  TVM_DEBUG_WITH_ABI_CHANGE: OFF
  USE_IOS_RPC: OFF
  USE_MSC: OFF
  USE_ETHOSU:
  CUDA_VERSION: NOT-FOUND
  USE_LIBBACKTRACE: AUTO
  DLPACK_PATH: 3rdparty/dlpack/include
  USE_TENSORRT_CODEGEN: OFF
  USE_THRUST: OFF
  USE_TARGET_ONNX: OFF
  USE_AOT_EXECUTOR: ON
  BUILD_DUMMY_LIBTVM: OFF
  USE_CUDNN: OFF
  USE_TENSORRT_RUNTIME: OFF
  USE_ARM_COMPUTE_LIB_GRAPH_EXECUTOR: OFF
  USE_CCACHE: AUTO
  USE_ARM_COMPUTE_LIB: OFF
  USE_CPP_RTVM:
  USE_OPENCL_GTEST: /path/to/opencl/gtest
  TVM_LOG_BEFORE_THROW: OFF
  USE_MKL: OFF
  USE_PT_TVMDSOOP: OFF
  MLIR_VERSION: NOT-FOUND
  USE_CLML: OFF
  USE_STACKVM_RUNTIME: OFF
  USE_GRAPH_EXECUTOR_CUDA_GRAPH: OFF
  ROCM_PATH: /opt/rocm
  USE_DNNL: OFF
  USE_MSCCL: OFF
  USE_VITIS_AI: OFF
  USE_MLIR: OFF
  USE_RCCL: OFF
  USE_LLVM: llvm-config --link-static
  USE_VERILATOR: OFF
  USE_TF_TVMDSOOP: OFF
  USE_THREADS: ON
  USE_MSVC_MT: OFF
  BACKTRACE_ON_SEGFAULT: OFF
  USE_GRAPH_EXECUTOR: ON
  USE_NCCL: OFF
  USE_ROCBLAS: OFF
  GIT_COMMIT_HASH: 2685d6ace64c30a077c1b3f6893d2e38589be7bb
  USE_VULKAN: OFF
  USE_RUST_EXT: OFF
  USE_CUTLASS: OFF
  USE_CPP_RPC: OFF
  USE_HEXAGON: OFF
  USE_CUSTOM_LOGGING: OFF
  USE_UMA: OFF
  USE_FALLBACK_STL_MAP: OFF
  USE_SORT: ON
  USE_RTTI: ON
  GIT_COMMIT_TIME: 2024-09-07 15:18:06 -0400
  USE_HIPBLAS: OFF
  USE_HEXAGON_SDK: /path/to/sdk
  USE_BLAS: none
  USE_ETHOSN: OFF
  USE_LIBTORCH: OFF
  USE_RANDOM: ON
  USE_CUDA: OFF
  USE_COREML: OFF
  USE_AMX: OFF
  BUILD_STATIC_RUNTIME: OFF
  USE_CMSISNN: OFF
  USE_KHRONOS_SPIRV: OFF
  USE_CLML_GRAPH_EXECUTOR: OFF
  USE_TFLITE: OFF
  USE_HEXAGON_GTEST: /path/to/hexagon/gtest
  PICOJSON_PATH: 3rdparty/picojson
  USE_OPENCL_ENABLE_HOST_PTR: OFF
  INSTALL_DEV: OFF
  USE_PROFILER: ON
  USE_NNPACK: OFF
  LLVM_VERSION: 17.0.1
  USE_MRVL: OFF
  USE_OPENCL: OFF
  COMPILER_RT_PATH: 3rdparty/compiler-rt
  RANG_PATH: 3rdparty/rang/include
  USE_SPIRV_KHR_INTEGER_DOT_PRODUCT: OFF
  USE_OPENMP: OFF
  USE_BNNS: OFF
  USE_FLASHINFER:
  USE_CUBLAS: OFF
  USE_METAL: ON
  USE_MICRO_STANDALONE_RUNTIME: OFF
  USE_HEXAGON_EXTERNAL_LIBS: OFF
  USE_ALTERNATIVE_LINKER: AUTO
  USE_BYODT_POSIT: OFF
  USE_NVSHMEM: OFF
  USE_HEXAGON_RPC: OFF
  USE_MICRO: OFF
  DMLC_PATH: 3rdparty/dmlc-core/include
  INDEX_DEFAULT_I64: ON
  USE_RELAY_DEBUG: OFF
  USE_RPC: ON
  USE_TENSORFLOW_PATH: none
  TVM_CLML_VERSION:
  USE_MIOPEN: OFF
  USE_ROCM: OFF
  USE_PAPI: OFF
  USE_CURAND: OFF
  TVM_CXX_COMPILER_PATH: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin/c++
  HIDE_PRIVATE_SYMBOLS: ON"
• Any other relevant information: I tried it with both the Mistral Small model from mlc-ai HF-repo, and also with a locally quantized version. Both yield the same error.

Additional context

Both models work fine separately.

[MasterJH5574]

Thank you @vlbosch. We also ran into this and get it fixed in #2906. The nightly packages are under built and will be ready in a few hours. I'll report back when the nightly build is done.

[MasterJH5574]

Hi @vlbosch the nightly wheel is updated and could you please try upgrade?

[vlbosch]

@MasterJH5574 Thanks for the quick response! I just updated to the latest nightly and retried. Small draft-mode does work now, however the speed running with small draft is slower than with Mistral Large alone. I thought that the baseline would be the regular speed of the large model? Or does that only count for the other speculative modes like eagle and medusa?

[MasterJH5574]

Hi @vlbosch thanks for following up and sorry for the late response. We have not yet benchmarked the speculative decoding performance on Metal so it's possible that the spec decoding for Metal needs more optimizations. It also depends on the speculative decoding acceptance of the 7B model.