[Model] support MiniMax-VL-01 model (#16328)

Signed-off-by: qingjun <[email protected]>

Author: qscqesze

tests/models/decoder_only/vision_language/test_models.py

@@ -446,6 +446,19 @@
         hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
         patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
     ),
+    "minimax_vl_01": VLMTestInfo(
+        models=["MiniMaxAI/MiniMax-VL-01"],
+        prompt_formatter=lambda img_prompt: f"<beginning_of_sentence>user: {img_prompt} assistant:<end_of_sentence>", # noqa: E501
+        img_idx_to_prompt=lambda _: "<image>",
+        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
+        max_model_len=8192,
+        max_num_seqs=4,
+        dtype="bfloat16",
+        hf_output_post_proc=model_utils.minimax_vl_01_hf_output,
+        patch_hf_runner=model_utils.minimax_vl_01_patch_hf_runner,
+        auto_cls=AutoModelForImageTextToText,
+        marks=[large_gpu_mark(min_gb=80)],
+    ),
     "molmo": VLMTestInfo(
         models=["allenai/Molmo-7B-D-0924"],
         test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),

tests/models/decoder_only/vision_language/vlm_utils/model_utils.py

@@ -229,6 +229,14 @@ def minicpmv_trunc_hf_output(hf_output: RunnerOutput,
     return output_ids, output_str, out_logprobs
 
 
+def minimax_vl_01_hf_output(hf_output: RunnerOutput,
+                            model: str) -> RunnerOutput:
+    output_ids, output_str, out_logprobs = hf_output
+    if output_str.endswith("<end_of_sentence>"):
+        output_str = output_str.split("<end_of_sentence>")[0]
+    return output_ids, output_str, out_logprobs
+
+
 ####### Functions for converting image assets to embeddings
 def get_llava_embeddings(image_assets: _ImageAssets):
     return [asset.image_embeds for asset in image_assets]
@@ -627,6 +635,17 @@ def _generate(self, *args, image_sizes=None, **kwargs):
     return hf_model
 
 
+def minimax_vl_01_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
+    orig_generate = hf_model.model.generate
+
+    def _generate(self, *args, image_sizes=None, **kwargs):
+        return orig_generate(*args, decode_text=False, **kwargs)
+
+    hf_model.model.generate = types.MethodType(_generate, hf_model.model)
+
+    return hf_model
+
+
 def molmo_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     """Patches and returns an instance of the HfRunner to use for Molmo."""
     hf_processor = hf_model.processor

tests/models/multimodal/processing/test_minimax_vl_01.py

@@ -0,0 +1,99 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import pytest
+from PIL import Image
+
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.parse import ImageSize
+from vllm.multimodal.processing import BaseMultiModalProcessor
+
+from ....conftest import _ImageAssets
+from ...utils import build_model_context
+
+
+@pytest.mark.parametrize("model_id", ["MiniMaxAI/MiniMax-VL-01"])
+# yapf: enable
+@pytest.mark.parametrize("num_imgs", [1, 2])
+def test_processor_override(
+    image_assets: _ImageAssets,
+    model_id: str,
+    num_imgs: int,
+):
+    ctx = build_model_context(
+        model_id,
+        mm_processor_kwargs=None,
+        limit_mm_per_prompt={"image": num_imgs},
+    )
+    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
+    prompt = "<image>" * num_imgs
+    image = Image.new("RGB", size=(364, 364))
+    mm_data = {"image": [image] * num_imgs}
+
+    processed_inputs = processor.apply(prompt, mm_data, {})
+    image_placeholders = processed_inputs["mm_placeholders"]["image"]
+
+    assert len(image_placeholders) == num_imgs
+
+
+def _validate_image_prompt_replacements_one(
+    processor: BaseMultiModalProcessor,
+    num_imgs: int,
+    failed_size_excs: list[tuple[ImageSize, Exception]],
+    image_size: ImageSize,
+) -> None:
+    prompt = "<image>" * num_imgs
+    image = Image.new("RGB", size=image_size)
+    mm_data = {"image": [image] * num_imgs}
+
+    try:
+        processed_inputs = processor.apply(prompt, mm_data, {})
+
+        image_placeholders = processed_inputs["mm_placeholders"]["image"]
+        assert len(image_placeholders) == num_imgs
+
+    except Exception as exc:
+        failed_size_excs.append((image_size, exc))
+
+
+def _test_image_prompt_replacements(
+    processor,
+    *,
+    num_imgs: int,
+    image_sizes: list[ImageSize],
+) -> None:
+
+    failed_size_excs = list[tuple[ImageSize, Exception]]()
+
+    for size in image_sizes:
+        _validate_image_prompt_replacements_one(processor, num_imgs,
+                                                failed_size_excs, size)
+
+    if failed_size_excs:
+        msg = "Found failing image sizes:" \
+            + "\n========\n".join(f"[{size}]\n{exc}"
+                                  for size, exc in failed_size_excs)
+        raise AssertionError(msg)
+
+
+@pytest.mark.parametrize("model_id", ["MiniMaxAI/MiniMax-VL-01"])
+@pytest.mark.parametrize("num_imgs", [1, 2])
+def test_processor_prompt_replacements_regression(model_id, num_imgs):
+    ctx = build_model_context(
+        model_id,
+        mm_processor_kwargs=None,
+        limit_mm_per_prompt={"image": num_imgs},
+    )
+    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
+
+    image_ratios = [(171, 152), (184, 161), (198, 176), (333, 296), (369, 328),
+                    (488, 183), (2560, 1669)]
+    image_sizes = [
+        size for w, h in image_ratios
+        for size in [ImageSize(w, h), ImageSize(h, w)]
+    ]
+
+    _test_image_prompt_replacements(
+        processor,
+        num_imgs=num_imgs,
+        image_sizes=image_sizes,
+    )

tests/models/registry.py

@@ -337,6 +337,8 @@ def check_available_online(
     "MiniCPMV": _HfExamplesInfo("openbmb/MiniCPM-Llama3-V-2_5",
                                 extras={"2.6": "openbmb/MiniCPM-V-2_6"},  # noqa: E501
                                 trust_remote_code=True),
+    "MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo("MiniMaxAI/MiniMax-VL-01", # noqa: E501
+                                              trust_remote_code=True),
     "Mistral3ForConditionalGeneration": _HfExamplesInfo("mistralai/Mistral-Small-3.1-24B-Instruct-2503",  # noqa: E501
                                                         extras={"fp8": "nm-testing/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic"}),  # noqa: E501
     "MolmoForCausalLM": _HfExamplesInfo("allenai/Molmo-7B-D-0924",

vllm/model_executor/models/minimax_text_01.py

@@ -3,7 +3,7 @@
 import copy
 import math
 import re
-from typing import Dict, Iterable, List, Optional, Tuple, Union
+from typing import Dict, Iterable, List, Optional, Set, Tuple, Union
 
 import torch
 import torch.distributed
@@ -110,7 +110,17 @@ def _forward(
             variance = tensor_model_parallel_all_reduce(
                 variance) / self.tp_world
         x = x * torch.rsqrt(variance + self.variance_epsilon)
-        x = x.to(orig_dtype) * self.weight
+
+        weight = self.weight
+        if x.size(-1) != self.weight.size(0):
+            if self.weight.size(0) < x.size(-1):
+                repeat_count = (x.size(-1) + self.weight.size(0)) // x.size(-1)
+                full_weight = self.weight.repeat(repeat_count)
+                weight = full_weight[:x.size(-1)]
+            else:
+                weight = self.weight[:x.size(-1)]
+
+        x = x.to(orig_dtype) * weight
         return x
 
     def forward(
@@ -421,6 +431,10 @@ def _prefill_and_mix_infer(self, q, k, v, kv_cache, state_indices_tensor,
                                attn_metadata):
         hidden = []
         for _prefill_idx in range(getattr(attn_metadata, "num_prefills", 0)):
+            if _prefill_idx >= len(attn_metadata.query_start_loc):
+                break
+            if _prefill_idx >= len(state_indices_tensor):
+                break
             _start = attn_metadata.query_start_loc[_prefill_idx]
             _end = attn_metadata.query_start_loc[_prefill_idx + 1]
             slot_id = state_indices_tensor[_prefill_idx]
@@ -443,6 +457,10 @@ def _prefill_and_mix_infer(self, q, k, v, kv_cache, state_indices_tensor,
             hidden.append(
                 self._decode_infer(q, k, v, kv_cache, state_indices_tensor,
                                    attn_metadata))
+
+        if not hidden:
+            return torch.empty((0, q.size(-1)), device=q.device, dtype=q.dtype)
+
         hidden = torch.concat(hidden, dim=0).contiguous()
         return hidden
 
@@ -663,6 +681,9 @@ def __init__(
         self.shared_moe = False
 
         shared_intermediate = getattr(config, 'shared_intermediate_size', 0)
+        if isinstance(shared_intermediate, list):
+            shared_intermediate = shared_intermediate[
+                layer_id] if layer_id < len(shared_intermediate) else 0
         if shared_intermediate > 0:
             self.shared_moe = True
             self.shared_mlp = MiniMaxText01MLP(
@@ -875,6 +896,8 @@ def _clear_prefill_cache(self, attn_metadata,
 
         slots_to_clear = []
         for _prefill_id in range(getattr(attn_metadata, "num_prefills", 0)):
+            if _prefill_id >= len(seq_id_map):
+                break
             seq_id = seq_id_map[_prefill_id]
             if attn_metadata.context_lens_tensor[
                     _prefill_id] == 0 and seq_id in seq_to_slot_maps:
@@ -886,13 +909,18 @@ def _clear_prefill_cache(self, attn_metadata,
                                         dtype=torch.long)
             minimax_cache_tensors[:, slots_tensor, ...] = 0
 
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+    ) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
     def forward(self,
                 input_ids: Optional[torch.Tensor],
                 positions: torch.Tensor,
-                kv_caches: List[torch.Tensor],
-                intermediate_tensors=None,
+                intermediate_tensors: Optional[IntermediateTensors] = None,
                 inputs_embeds: Optional[torch.Tensor] = None,
-                **kwargs) -> torch.Tensor:
+                **kwargs) -> Union[torch.Tensor, IntermediateTensors]:
         forward_context = get_forward_context()
         attn_metadata = forward_context.attn_metadata
         if attn_metadata is None:
@@ -901,6 +929,7 @@ def forward(self,
             kwargs["request_ids_to_seq_ids"] = {}
         if "finished_requests_ids" not in kwargs:
             kwargs["finished_requests_ids"] = []
+
         (
             minimax_cache_tensors,
             state_indices_tensor,
@@ -922,15 +951,11 @@ def forward(self,
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        kv_cache_index = 0
         minimax_cache_index = 0
         attn_metadata.rotary_emb = self.rotary_emb
         for i in range(self.start_layer, self.end_layer):
             layer = self.layers[i]
             _caches = None
-            if isinstance(layer.self_attn, MiniMaxText01Attention):
-                _caches = kv_caches[kv_cache_index]
-                kv_cache_index += 1
             if isinstance(layer.self_attn, MiniMaxText01LinearAttention):
                 current_state_layer = minimax_cache_index
                 _caches = minimax_cache_params.at_layer_idx(
@@ -1009,15 +1034,20 @@ def get_seqlen_agnostic_capture_inputs(self, batch_size: int):
         return self.model.minimax_cache.get_seqlen_agnostic_capture_inputs(
             batch_size)
 
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+    ) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
     def forward(self,
                 input_ids: torch.Tensor,
                 positions: torch.Tensor,
                 intermediate_tensors: Optional[IntermediateTensors] = None,
                 inputs_embeds: Optional[torch.Tensor] = None,
                 **kwargs) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, self.kv_cache,
-                                   intermediate_tensors, inputs_embeds,
-                                   **kwargs)
+        hidden_states = self.model(input_ids, positions, intermediate_tensors,
+                                   inputs_embeds, **kwargs)
 
         return hidden_states
 
@@ -1043,8 +1073,9 @@ def make_empty_intermediate_tensors(
         })
 
     def load_weights(self, weights: Iterable[Tuple[str,
-                                                   torch.Tensor]]) -> None:
+                                                   torch.Tensor]]) -> Set[str]:
         params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
 
         def which_layer(name: str) -> int:
             if "layers" in name:
@@ -1108,6 +1139,7 @@ def load_sparse_moe_weight(name: str, loaded_weight: torch.Tensor,
                               weight_name,
                               expert_id=expert_id,
                               shard_id=shard_id)
+                loaded_params.add(name)
                 break
             else:
                 if is_pp_missing_parameter(name, self):
@@ -1117,6 +1149,7 @@ def load_sparse_moe_weight(name: str, loaded_weight: torch.Tensor,
                                         default_weight_loader)
                 weight_loader = weight_loader_with_alias(name)(weight_loader)
                 weight_loader(param, loaded_weight)
+                loaded_params.add(name)
             return
 
         def is_shared_mlp_weight(name: str) -> bool:
@@ -1154,6 +1187,7 @@ def load_shared_mlp_weight(name: str, loaded_weight: torch.Tensor,
                 else:
                     raise AssertionError(
                         "MLP weight not in [gate_up_proj, down_proj]")
+            loaded_params.add(name)
             return
 
         def is_mha_weight(name: str) -> bool:
@@ -1170,6 +1204,7 @@ def load_linear_attn_weight(name: str, loaded_weight: torch.Tensor,
                 MiniMaxText01LinearAttention.weight_direct_load)
             weight_loader = weight_loader_with_alias(name)(weight_loader)
             weight_loader(param, loaded_weight)
+            loaded_params.add(name)
             return
 
         def load_flash_attn_weight(name: str, loaded_weight: torch.Tensor,
@@ -1194,6 +1229,7 @@ def load_flash_attn_weight(name: str, loaded_weight: torch.Tensor,
                                         default_weight_loader)
                 weight_loader = weight_loader_with_alias(name)(weight_loader)
                 weight_loader(param, loaded_weight, shard_id)
+                loaded_params.add(name)
                 break
             else:
                 if is_pp_missing_parameter(name, self):
@@ -1204,6 +1240,7 @@ def load_flash_attn_weight(name: str, loaded_weight: torch.Tensor,
                                         default_weight_loader)
                 weight_loader = weight_loader_with_alias(name)(weight_loader)
                 weight_loader(param, loaded_weight)
+                loaded_params.add(name)
             return
 
         def is_layer_norm_weight(name: str) -> bool:
@@ -1219,6 +1256,7 @@ def load_layer_norm_weight(name: str, loaded_weight: torch.Tensor,
                                     default_weight_loader)
             weight_loader = weight_loader_with_alias(name)(weight_loader)
             weight_loader(param, loaded_weight)
+            loaded_params.add(name)
             return
 
         def load_basic_weight(name: str, loaded_weight: torch.Tensor,
@@ -1230,6 +1268,7 @@ def load_basic_weight(name: str, loaded_weight: torch.Tensor,
                                     default_weight_loader)
             weight_loader = weight_loader_with_alias(name)(weight_loader)
             weight_loader(param, loaded_weight)
+            loaded_params.add(name)
             return
 
         for name, loaded_weight in weights:
@@ -1258,4 +1297,4 @@ def load_basic_weight(name: str, loaded_weight: torch.Tensor,
                 continue
 
             load_basic_weight(name, loaded_weight, self)
-        return
+        return loaded_params