86d6c9f6c2
FOUND: baseline.muh had completely different values from computility-run.yaml (the actual deployment config). This means gen_yaml.py would produce a WRONG computility-run.yaml if someone regenerated it from baseline.muh. Key differences synced: max_model_len: 100000 → 256000 (competition allows 256K context) gpu_memory_utilization: 0.9 → 0.95 (squeeze more KV cache) max_num_seqs: 1 → 2 (allow 2 concurrent sequences) max_num_batched_tokens: 8192 → 4096 (smaller prefill chunks) enforce_eager: (missing) → true (BI-V100 doesn't support CUDA graph) dtype: (missing) → half VLLM_ATTENTION_BACKEND: (missing) → XFORMERS CRITICAL DISCOVERY: CoreX native libraries revealed: libcorex_fa2.so — Iluvatar FlashAttention2 (NOT generic xformers) libcorex_gdn.so — CoreX GDN ops libcorex_moe.so — CoreX MoE GEMM kernel These are the REAL performance-critical kernels, loaded via VLLM_COREX_* env vars. The Triton flash_attention.py is a FALLBACK, not the primary path. CCCL insight: thread_store.cuh shows PTX cache modifiers (st.cg, st.cs) may be ignored on non-NVIDIA hardware. This explains why LOAD_DEFAULT outperforms LOAD_LDG on BI-V100 — CoreX has a different cache hierarchy.
49 lines
1.6 KiB
Plaintext
49 lines
1.6 KiB
Plaintext
# baseline.muh — Competition vllm launch configuration
|
|
# SYNCED FROM computility-run.yaml (the actual deployment config)
|
|
#
|
|
# This file stores ONLY the vllm server launch config.
|
|
# Kernel tuning values live in muh/include/muh/tuning/tuning_*.cuh
|
|
# as constexpr structs — NOT here.
|
|
#
|
|
# Pipeline:
|
|
# muh/tuning/*.cuh (bi100_* values) → gen_patch.py → vllm kernel patches
|
|
# baseline.muh (vllm config) → gen_yaml.py → computility-run.yaml
|
|
#
|
|
# CRITICAL: computility-run.yaml is the deployment source of truth.
|
|
# This .muh must stay in sync with it.
|
|
|
|
# --- vllm launch configuration ---
|
|
vllm:
|
|
model_path: /model
|
|
served_model_name: llm
|
|
max_model_len: 256000
|
|
gpu_memory_utilization: 0.95
|
|
tensor_parallel: 4
|
|
max_num_seqs: 2
|
|
max_num_batched_tokens: 4096
|
|
max_seq_len_to_capture: 32768
|
|
trust_remote_code: true
|
|
disable_log_requests: true
|
|
disable_frontend_multiprocessing: true
|
|
enable_chunked_prefill: true
|
|
enable_auto_tool_choice: true
|
|
tool_call_parser: qwen3_coder
|
|
reasoning_parser: qwen3
|
|
enable_prefix_caching: true
|
|
enforce_eager: true
|
|
dtype: half
|
|
|
|
concurrency: 1
|
|
|
|
env:
|
|
VLLM_ENGINE_ITERATION_TIMEOUT_S: 3600
|
|
VLLM_ATTENTION_BACKEND: XFORMERS
|
|
ENABLE_CUSTOM_IPC: 1
|
|
PYTHONPATH: /usr/local/corex/lib/python3/dist-packages:/usr/local/corex/lib64/python3/dist-packages
|
|
LD_LIBRARY_PATH: /usr/local/corex/lib64:/usr/local/openmpi/lib
|
|
VLLM_COREX_FA2_LIBRARY: /usr/local/corex/lib64/libcorex_fa2.so
|
|
VLLM_COREX_GDN_LIBRARY: /usr/local/corex/lib64/libcorex_gdn.so
|
|
VLLM_COREX_MOE_LIBRARY: /usr/local/corex/lib64/libcorex_moe.so
|
|
VLLM_REQUEST_METRICS_FILE: /tmp/vllm-request-metrics.jsonl
|
|
VLLM_CACHE_BLOCK_SIZE: 16
|