[AMD] MiniMax-M3: opt-in custom/quick all-reduce on ROCm (#32230)

python/sglang/srt/arg_groups/overrides.py

@@ -508,7 +508,12 @@ def _minimax_m3_overrides(server_args: Any, hf_config: Any) -> dict:
             )
             overrides["enable_aiter_allreduce_fusion"] = False
             aiter_fusion_resolved = False
-        if not aiter_fusion_resolved:
+        # By default MiniMax-M3 on ROCm keeps NCCL all-reduce (custom AR off)
+        # whenever aiter all-reduce fusion is not used. Opting in via
+        # SGLANG_M3_ALLOW_CUSTOM_AR keeps custom all-reduce enabled so the
+        # quick-reduce path (ROCM_QUICK_REDUCE_QUANTIZATION=INT4/INT6/INT8) can
+        # accelerate the large prefill all-reduce.
+        if not aiter_fusion_resolved and not envs.SGLANG_M3_ALLOW_CUSTOM_AR.get():
             overrides["disable_custom_all_reduce"] = True
     elif is_sm100_supported():
         if server_args.is_attention_backend_not_set():

python/sglang/srt/environ.py

@@ -815,6 +815,11 @@ class Envs:
     # Set to 0: force disable (use default Aiter AR even with --enable-deterministic-inference)
     SGLANG_USE_1STAGE_ALLREDUCE = EnvBool(False)
     SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2 = EnvBool(True)
+    # MiniMax-M3 on ROCm force-disables custom all-reduce in its model override
+    # (arg_groups/overrides.py) when aiter all-reduce fusion is off. Set this to
+    # opt back in and keep custom/quick all-reduce enabled -- e.g. to run the
+    # INT4 quick-reduce path via ROCM_QUICK_REDUCE_QUANTIZATION={INT4,INT6,INT8}.
+    SGLANG_M3_ALLOW_CUSTOM_AR = EnvBool(False)
     # Default per-direction workspace cap for CustomAllReduceV2; explicit
     # constructor sizes take precedence over this.
     SGLANG_CUSTOM_ALL_REDUCE_V2_MAX_SIZE_KB = EnvInt(16 * 1024)