[GLM-5] Apply trtllm MHA kernel for GLM-5 on Blackwell (#21332)

2026-06-03 17:16:54 -07:00
parent 3790173b3b
commit d097cd2212
1 changed files with 10 additions and 16 deletions
@@ -1831,12 +1831,6 @@ class ServerArgs:
        ]:
            # Set attention backend for DeepSeek
            if is_deepseek_dsa(hf_config):  # DeepSeek 3.2/GLM 5
-                if model_arch == "GlmMoeDsaForCausalLM" and is_blackwell_supported():
-                    envs.SGLANG_DSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD.set(0)
-                    logger.warning(
-                        "Force DSA prefill to use sparse MLA (i.e. disable MHA_ONE_SHOT) for GlmMoeDsaForCausalLM on Blackwell."
-                    )
-                else:
                if envs.SGLANG_DSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD.is_set():
                    logger.warning(
                        f"Dense attention kv len threshold is manually set to {envs.SGLANG_DSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD.get()} for DSA. Caution: This may cause performance regression if the threshold is larger than the index topk of model."