[RL] Refactor NVFP4 shuffling/swizzling to in-place replacement (#22204)

2026-04-12 19:08:45 -07:00
parent 934e19a610
commit 5593539942
6 changed files with 157 additions and 103 deletions
@@ -2824,8 +2824,9 @@ class ServerArgs:
            assert self.quantization in [
                "fp8",
                "mxfp8",
+                "modelopt_fp4",
                None,
-            ], f"Invalid quantization '{self.quantization}'. \nFlashInfer TRTLLM routed MOE supports only: 'fp8', 'mxfp8', or bfloat16 (None)."
+            ], f"Invalid quantization '{self.quantization}'. \nFlashInfer TRTLLM routed MOE supports only: 'fp8', 'mxfp8', 'modelopt_fp4', or bfloat16 (None)."
            self.disable_shared_experts_fusion = True
            logger.warning(
                "FlashInfer TRTLLM routed MoE is enabled. --disable-shared-experts-fusion is automatically set."