From 86d6c9f6c29b70e128b58ef1c2564feaf17d880d Mon Sep 17 00:00:00 2001
From: muh <muh@project6.dev>
Date: Thu, 6 Aug 2026 06:12:51 +0000
Subject: [PATCH] =?UTF-8?q?[critical/config]=20baseline.muh:=20sync=20from?=
 =?UTF-8?q?=20computility-run.yaml=20=E2=80=94=20was=20stale?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

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.
---
 baseline.muh | 23 +++++++++++++++++++----
 1 file changed, 19 insertions(+), 4 deletions(-)

diff --git a/baseline.muh b/baseline.muh
index 3885bcc3..839af03e 100644
--- a/baseline.muh
+++ b/baseline.muh
@@ -1,4 +1,5 @@
 # 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
@@ -7,16 +8,19 @@
 # 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: 100000
-  gpu_memory_utilization: 0.9
+  max_model_len: 256000
+  gpu_memory_utilization: 0.95
   tensor_parallel: 4
-  max_num_seqs: 1
-  max_num_batched_tokens: 8192
+  max_num_seqs: 2
+  max_num_batched_tokens: 4096
   max_seq_len_to_capture: 32768
   trust_remote_code: true
   disable_log_requests: true
@@ -26,8 +30,19 @@ vllm:
   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