| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768 |
- engine = "vllm"
- name = "DeepSeek V4 Flash (vLLM)"
- docker_image = "aidendle94/sparkrun-vllm-ds4-gb10:production-v2"
- model = "deepseek-ai/DeepSeek-V4-Flash"
- required_nodes = 2
- tensor_parallel_size = 2
- pipeline_parallel_size = 1
- distributed_backend = "mp"
- # DeepSeek V4 with 1M context needs much more shared memory than the
- # default 10g. KV cache + activations for 1M tokens fills up fast.
- shm_size = "64gb"
- entrypoint = ["vllm", "serve"]
- args = [
- "--kv-cache-dtype", "fp8",
- "--block-size", "256",
- "--max-model-len", "524288",
- "--max-num-seqs", "8",
- "--max-num-batched-tokens", "8192",
- "--enable-chunked-prefill",
- "--gpu-memory-utilization", "0.85",
- "--compilation-config", """
- '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'""",
- "--enable-prefix-caching",
- "--load-format", "safetensors",
- "--speculative-config", """
- '{"method":"mtp","num_speculative_tokens":2}'""",
- "--tokenizer-mode", "deepseek_v4",
- "--tool-call-parser", "deepseek_v4",
- "--enable-auto-tool-choice",
- "--reasoning-parser", "deepseek_v4",
- "--reasoning-config", """
- '{"reasoning_parser":"deepseek_v4","reasoning_start_str":"","reasoning_end_str":""}'""",
- "--default-chat-template-kwargs", """
- '{"thinking":true,"preserve_thinking":true}'""",
- "--default-chat-template-kwargs.reasoning_effort", "high",
- "--enable-flashinfer-autotune",
- "--trust-remote-code",
- "--served-model-name", "deepseek-v4-flash", "default",
- ]
- # --distributed-executor-backend mp
- [env]
- # Ray memory management — prevents OOM during cudagraph compilation
- RAY_memory_monitor_refresh_ms = "0"
- RAY_num_prestart_python_workers = "0"
- RAY_object_store_memory = "1073741824"
- # vLLM features for DeepSeek V4 on Blackwell (GB10 / sm_120)
- VLLM_ALLOW_LONG_MAX_MODEL_LEN = "1"
- VLLM_USE_B12X_MOE = "1"
- VLLM_SPARSE_INDEXER_MAX_LOGITS_MB = "256"
- # Compile-time arch hints for PyTorch and FlashInfer on GB10
- TORCH_CUDA_ARCH_LIST = "12.1a"
- FLASHINFER_CUDA_ARCH_LIST = "12.1a"
- CUDA_HOME = "/opt/env/targets/sbsa-linux"
- FLASHINFER_DISABLE_VERSION_CHECK = "1"
- TILELANG_CLEANUP_TEMP_FILES = "1"
- DG_JIT_USE_NVRTC = "0"
- DG_JIT_PRINT_COMPILER_COMMAND = "1"
- VLLM_NCCL_SO_PATH = "/opt/env/lib/python3.12/site-packages/nvidia/nccl/lib/libnccl.so.2"
- HF_HUB_OFFLINE = "1"
- TRANSFORMERS_OFFLINE = "1"
|
