[ROCm][Quant] Minmax-M3: enable fp8_per_channel and fix SwiGLU-OAI fp8 MoE for bf16 weights on mi300x

vllm/model_executor/layers/fused_moe/config.py

@@ -603,10 +603,17 @@ def fp8_w8a8_moe_quant_config(
     a2_gscale: torch.Tensor | None = None,
     g1_alphas: torch.Tensor | None = None,
     g2_alphas: torch.Tensor | None = None,
+    gemm1_alpha: float | None = None,
+    gemm1_beta: float | None = None,
     gemm1_clamp_limit: float | None = None,
 ) -> FusedMoEQuantConfig:
     """
     Construct a quant config for fp8 activations and fp8 weights.
+
+    ``gemm1_alpha``/``gemm1_beta`` drive the SwiGLU-OAI GEMM1 activation
+    (``silu_and_mul_with_clamp``); they must be forwarded for SwiGLU-OAI MoEs
+    (e.g. MiniMax-M3 alpha=1.702, beta=1.0), else the kernel defaults to
+    alpha=1.0/beta=0.0 and produces garbage.
     """
     return FusedMoEQuantConfig.make(
         current_platform.fp8_dtype(),
@@ -623,6 +630,8 @@ def fp8_w8a8_moe_quant_config(
         per_act_token_quant=per_act_token_quant,
         per_out_ch_quant=per_out_ch_quant,
         block_shape=block_shape,
+        gemm1_alpha=gemm1_alpha,
+        gemm1_beta=gemm1_beta,
         gemm1_clamp_limit=gemm1_clamp_limit,
     )
 

vllm/model_executor/layers/fused_moe/oracle/fp8.py

@@ -568,6 +568,8 @@ def make_fp8_moe_quant_config(
         block_shape=block_shape,
         per_act_token_quant=per_act_token_quant,
         per_out_ch_quant=per_out_ch_quant,
+        gemm1_alpha=gemm1_alpha,
+        gemm1_beta=gemm1_beta,
         gemm1_clamp_limit=swiglu_limit,
     )
 

vllm/model_executor/layers/quantization/fp8.py

@@ -849,6 +849,11 @@ def get_fused_moe_quant_config(self, layer: RoutedExperts) -> FusedMoEQuantConfi
             a2_scale=a2_scale,
             block_shape=self.weight_block_size,
             swiglu_limit=getattr(layer, "swiglu_limit", None),
+            # SwiGLU-OAI alpha/beta (e.g. MiniMax-M3: 1.702/1.0). Without these
+            # the fp8 MoE applies silu_and_mul_with_clamp with the default
+            # alpha=1.0/beta=0.0, producing garbage for SwiGLU-OAI MoEs.
+            gemm1_alpha=getattr(layer, "swiglu_alpha", None),
+            gemm1_beta=getattr(layer, "swiglu_beta", None),
         )
 
         # Inject biases into the quant config if the model has them

vllm/model_executor/layers/quantization/online/fp8.py

@@ -476,6 +476,11 @@ def get_fused_moe_quant_config(
             per_act_token_quant=self.per_act_token_quant,
             per_out_ch_quant=self.per_out_ch_quant,
             swiglu_limit=getattr(layer, "swiglu_limit", None),
+            # SwiGLU-OAI alpha/beta (e.g. MiniMax-M3: 1.702/1.0). Without these
+            # the online fp8 MoE runs silu_and_mul_with_clamp with the default
+            # alpha=1.0/beta=0.0, producing garbage for SwiGLU-OAI MoEs.
+            gemm1_alpha=getattr(layer, "swiglu_alpha", None),
+            gemm1_beta=getattr(layer, "swiglu_beta", None),
         )
 
 

vllm/platforms/rocm.py

@@ -459,6 +459,7 @@ class RocmPlatform(Platform):
         "modelopt_mixed",
         "fp8_per_tensor",
         "fp8_per_block",
+        "fp8_per_channel",  # PTPC: per-channel weight + per-token act fp8
         "online",
         "gpt_oss_mxfp4",
     ]