Use Flashinfer TRT-LLM as Llama 4 compatible MoE backend (#11928)

2025-10-28 10:39:43 -07:00
parent 9c6e25d2a6
commit 77225d602a
4 changed files with 171 additions and 7 deletions
@@ -643,6 +643,7 @@ class ModelConfig:
            "petit_nvfp4",
        ]
        compatible_quantization_methods = {
+            "modelopt_fp8": ["modelopt"],
            "modelopt_fp4": ["modelopt"],
            "petit_nvfp4": ["modelopt"],
            "w8a8_int8": ["compressed-tensors", "compressed_tensors"],