[Bugfix][Model] Fix Devstral Small 2 HF format weight loading (#39293)

Signed-off-by: thomasmaindron <thomasmaindron@users.noreply.github.com>
Co-authored-by: thomasmaindron <thomasmaindron@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

tests/models/registry.py

@@ -416,6 +416,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
         "MiniMaxAI/MiniMax-M2",
         trust_remote_code=True,
     ),
+    "Ministral3ForCausalLM": _HfExamplesInfo("mistralai/Ministral-3-3B-Instruct-2512"),
     "MistralForCausalLM": _HfExamplesInfo("mistralai/Mistral-7B-Instruct-v0.1"),
     "MistralLarge3ForCausalLM": _HfExamplesInfo(
         "mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4"

vllm/model_executor/models/mistral3.py

@@ -382,7 +382,14 @@ class Mistral3ForConditionalGeneration(
             # Some PEFT LoRAs are trained against the text submodule directly
             # and produce names like `base_model.model.model.layers.*`.
             "model.": "language_model.model.",
-        }
+        },
+        orig_to_new_suffix={
+            # FP8 quantized HF checkpoints use "activation_scale" and
+            # "weight_scale_inv" but vLLM's FP8 linear layers register
+            # them as "input_scale" and "weight_scale"
+            ".activation_scale": ".input_scale",
+            ".weight_scale_inv": ".weight_scale",
+        },
     )
 
     @classmethod
@@ -402,13 +409,8 @@ class Mistral3ForConditionalGeneration(
         self.config = config
         self.multimodal_config = multimodal_config
 
-        # NOTE: These are special cases for Pixtral-12B in the HF-format
+        # NOTE: This is a special case for Pixtral-12B in the HF-format
         # https://huggingface.co/mistral-community/pixtral-12b/blob/main/config.json  # noqa
-        if (
-            config.text_config.architectures is None
-            and config.text_config.model_type == "mistral"
-        ):
-            config.text_config.architectures = ["MistralForCausalLM"]
         if (
             config.projector_hidden_act is None
             and config.vision_config.hidden_act == "gelu"

vllm/model_executor/models/registry.py

@@ -160,6 +160,7 @@ _TEXT_GENERATION_MODELS = {
     "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
     "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
     "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
+    "Ministral3ForCausalLM": ("mistral", "MistralForCausalLM"),
     "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
     "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
     "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),