feat: [Qwen3.5] Support block-wise FP8 quantization and model adaptation (#18926)

2026-02-18 11:44:25 +08:00
parent 83e24e2eb4
commit fa5698d791
4 changed files with 57 additions and 12 deletions
@@ -171,8 +171,11 @@ class Fp8Config(QuantizationConfig):
            config, ["ignored_layers", "modules_to_not_convert"], None
        )
        if ignored_layers:
-            # hack for ministral
-            ignored_layers = [layer.replace("model.", "") for layer in ignored_layers]
+            if "mistral3" in config.get("model_type", ""):
+                # hack for ministral
+                ignored_layers = [
+                    layer.replace("model.", "") for layer in ignored_layers
+                ]
        weight_block_size = cls.get_from_keys_or(config, ["weight_block_size"], None)
        if use_mxfp8 and weight_block_size is not None:
            logger.warning(