|
|
@@ -69,7 +69,7 @@ args = [
|
|
|
"--reasoning-parser", "qwen3",
|
|
|
"--chat-template", "/workspace/flashnext/model/medium_chat_template.jinja",
|
|
|
"--speculative-config",
|
|
|
- '{"method":"mtp","num_speculative_tokens":3,"model":"/workspace/flashnext/draft-k10","rejection_sample_method":"block","draft_sample_method":"probabilistic"}',
|
|
|
+ '{"method":"mtp","num_speculative_tokens":4,"model":"/workspace/flashnext/draft-k10","rejection_sample_method":"block","draft_sample_method":"probabilistic"}',
|
|
|
]
|
|
|
|
|
|
[mods]
|
|
|
@@ -88,7 +88,7 @@ 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 = "0" # 2026-09-26: off — garble (653x171->"11") caught while a ~70K-token request co-batched; atomic path is m/n/k-shape dependent. A/B spare on dgx3 kept at 1.
|
|
|
+VLLM_MARLIN_USE_ATOMIC_ADD = "1" # 2026-09-26: back to 1 (the bench baseline). The 1→0 flip was a garble red herring — the real cause was the unpinned b12x image (see above).
|
|
|
VLLM_USE_DEEP_GEMM = "0"
|
|
|
VLLM_USE_FLASHINFER_SAMPLER = "1"
|
|
|
VLLM_MTP_DRAFT_SCALE = "2"
|
