MiniMax-M3: allow shared-experts fusion on ROCm gfx942 and newer (#36576)

Co-authored-by: Kevin Mi <kevin.mi@radixark.ai>

python/sglang/srt/layers/moe/topk.py

@@ -2404,9 +2404,10 @@ def select_experts(
     # slots on the marker) and places that marker at id num_experts, which the
     # DeepEP remap shifts one past the end of the expert space -- 384 -> 392 for
     # 384 routed experts on EP8, where the valid ids are 0..391.
+    # aiter appends the shared expert in _post_process_topk_ids, so the gate must not.
     num_fused_shared_experts_for_gate = (
         0
-        if has_per_rank_fused_shared_slots(num_fused_shared_experts)
+        if (has_per_rank_fused_shared_slots(num_fused_shared_experts) or _use_aiter)
         else num_fused_shared_experts
     )
     if dynamic_expert_bias is not None:

python/sglang/srt/models/minimax_m3.py

@@ -1612,10 +1612,12 @@ class MiniMaxM3SparseForCausalLM(nn.Module):
                 "Shared and routed experts may use different quantization formats "
                 "in ModelOpt mixed-precision checkpoints."
             )
-        if not _is_cuda:
-            return "Shared experts fusion currently requires CUDA devices."
+        if not (_is_cuda or _is_hip):
+            return "Shared experts fusion currently requires CUDA or ROCm devices."
         if _is_cuda and (_device_sm is not None) and (_device_sm < 80):
             return "Shared experts fusion requires SM80 or newer GPUs."
+        if _is_hip and not _is_gfx95_supported:
+            return "Shared experts fusion on ROCm is validated on gfx950 only."
         if get_parallel().moe_ep_size > 1:
             return "Shared experts fusion is not supported together with expert parallelism yet."
         if get_moe_a2a_backend().is_deepep():

python/sglang/srt/models/minimax_m3_vl.py

@@ -47,13 +47,22 @@ from sglang.srt.models.minimax_vl_common import (
 )
 from sglang.srt.models.utils import WeightsMapper
 from sglang.srt.runtime_context import get_mm, get_parallel
-from sglang.srt.utils import add_prefix, get_device_sm, is_cuda, log_info_on_rank0
+from sglang.srt.utils import (
+    add_prefix,
+    get_device_sm,
+    is_cuda,
+    is_gfx95_supported,
+    is_hip,
+    log_info_on_rank0,
+)
 from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 logger = logging.getLogger(__name__)
 
 
 _is_cuda = is_cuda()
+_is_hip = is_hip()
+_is_gfx95_supported = is_gfx95_supported()
 _device_sm = get_device_sm()
 
 
@@ -152,10 +161,12 @@ class MiniMaxM3SparseForConditionalGeneration(nn.Module):
                 "Shared and routed experts may use different quantization formats "
                 "in ModelOpt mixed-precision checkpoints."
             )
-        if not _is_cuda:
-            return "Shared experts fusion currently requires CUDA devices."
-        if (_device_sm is not None) and (_device_sm < 80):
+        if not (_is_cuda or _is_hip):
+            return "Shared experts fusion currently requires CUDA or ROCm devices."
+        if _is_cuda and (_device_sm is not None) and (_device_sm < 80):
             return "Shared experts fusion requires SM80 or newer GPUs."
+        if _is_hip and not _is_gfx95_supported:
+            return "Shared experts fusion on ROCm is validated on gfx950 only."
         if get_parallel().moe_ep_size > 1:
             return (
                 "Shared experts fusion is not supported together with expert "