Selaa lähdekoodia

qwen: VLLM_MARLIN_USE_ATOMIC_ADD off (experimental atomic-add path, suspected intermittent garble)

Alexander Korolev 3 viikkoa sitten
vanhempi
sitoutus
7250e2ad11
1 muutettua tiedostoa jossa 1 lisäystä ja 1 poistoa
  1. 1 1
      recipes/qwen38-flash-next-int4/recipe.toml

+ 1 - 1
recipes/qwen38-flash-next-int4/recipe.toml

@@ -83,7 +83,7 @@ B12X_POLICY_MODE = "auto"
 VLLM_PLE_TABLE_MEMORY = "disk"
 # this checkpoint (the mod patches them into the container's vLLM)
 VLLM_FP8_HYBRID = "1"
-VLLM_MARLIN_USE_ATOMIC_ADD = "1"
+VLLM_MARLIN_USE_ATOMIC_ADD = "0"  # was 1 (forum yaml); 1 = experimental atomic-add path, intermittent garbage
 VLLM_USE_DEEP_GEMM = "0"
 VLLM_USE_FLASHINFER_SAMPLER = "1"
 VLLM_MTP_DRAFT_SCALE = "2"