fix(profile): disable Flash vLLM torch compile

Oseltamivir · Oseltamivir · commit fc21e40e15cd · 2026-05-21T20:48:26.000-07:00
diff --git a/benchmarks/single_node/dsv4_fp4_b300_vllm_mtp.sh b/benchmarks/single_node/dsv4_fp4_b300_vllm_mtp.sh
@@ -61,7 +61,7 @@ COMPILATION_ARGS=(
     --max-cudagraph-capture-size 2048
 )
 if [[ "$MODEL" == "deepseek-ai/DeepSeek-V4-Flash" ]]; then
-    COMPILATION_ARGS=(--compilation-config '{"cudagraph_mode":"NONE","custom_ops":["all"]}')
+    COMPILATION_ARGS=(--compilation-config '{"mode":0,"cudagraph_mode":"NONE","custom_ops":["all"]}')
 fi
 
 BENCHMARK_MAX_MODEL_LEN=$MAX_MODEL_LEN

Original file line number	Diff line number	Diff line change
`@@ -61,7 +61,7 @@ COMPILATION_ARGS=(`
`61`	`61`	`--max-cudagraph-capture-size 2048`
`62`	`62`	`)`
`63`	`63`	`if [[ "$MODEL" == "deepseek-ai/DeepSeek-V4-Flash" ]]; then`
`64`		`- COMPILATION_ARGS=(--compilation-config '{"cudagraph_mode":"NONE","custom_ops":["all"]}')`
	`64`	`+ COMPILATION_ARGS=(--compilation-config '{"mode":0,"cudagraph_mode":"NONE","custom_ops":["all"]}')`
`65`	`65`	`fi`
`66`	`66`
`67`	`67`	`BENCHMARK_MAX_MODEL_LEN=$MAX_MODEL_LEN`