[Feature] add MiniMax M3 VL model support (vllm-project#12844)

Migrate the MiniMax M3 vision-language model implementation from
vllm-project#12734 into the synced main branch as part of
the split MiniMax M3 support.

### What this PR does / why we need it?

This PR adds `vllm_ascend/models/minimax_m3/minimax_m3_vl.py`, the
vision-language model wrapper for MiniMax M3 on Ascend.

This PR only contains the VL model file. The sparse attention ops, text
model, model registry, quantization, and worker-side changes are split
into separate PRs.

### Does this PR introduce _any_ user-facing change?

No standalone user-facing change.

This PR adds one implementation file as part of the split MiniMax M3
support. The model becomes user-visible only after the remaining MiniMax
M3 components and model registration are merged.

### How was this patch tested?

- Validated in the full MiniMax M3 integration branch from
vllm-project#12734.
- Service startup passed with TP=8 and EP enabled.
- Text, image, and video single-request smoke tests passed.
- A long-context text request with `usage.prompt_tokens = 3183` passed,
covering the >2048-token index-cache scenario discussed in review.

- vLLM version: v0.25.1
- vLLM main:
vllm-project/vllm@d02df74

Signed-off-by: muziyuhui666 <lijianfu9@huawei.com>
Co-authored-by: CXY-Katrina <katrina.cxy@gmail.com>
Co-authored-by: cywang250805 <wangchaoyu7@huawei.com>
Co-authored-by: Bill845514379 <huangjianbao2@huawei.com>
Co-authored-by: yejj710 <yejj710@gmail.com>
Co-authored-by: AuroraEmiya <Sakura.iostream@gmail.com>
Co-authored-by: HaoxinZong <116423146+HaoxinZong@users.noreply.github.com>

Author: 7 people

vllm_ascend/models/__init__.py

@@ -7,6 +7,10 @@ def register_model():
         "MiniMaxM3SparseForCausalLM",
         "vllm_ascend.models.minimax_m3:MiniMaxM3SparseForCausalLM",
     )
+    ModelRegistry.register_model(
+        "MiniMaxM3SparseForConditionalGeneration",
+        "vllm_ascend.models.minimax_m3:MiniMaxM3SparseForConditionalGeneration",
+    )
     ModelRegistry.register_model("DeepSeekV4MTPModel", "vllm_ascend.models.deepseek_v4_mtp:DeepSeekV4MTP")
     ModelRegistry.register_model(
         "DSparkDraftModel",

vllm_ascend/models/minimax_m3/__init__.py

@@ -8,12 +8,14 @@
     _get_rope_parameters,
     _sparse_attention_layer_ids,
 )
+from vllm_ascend.models.minimax_m3.minimax_m3_vl import MiniMaxM3SparseForConditionalGeneration
 
 __all__ = [
     "MiniMaxM3Attention",
     "MiniMaxM3MoE",
     "MiniMaxM3SparseAttention",
     "MiniMaxM3SparseForCausalLM",
+    "MiniMaxM3SparseForConditionalGeneration",
     "_get_rope_parameters",
     "_sparse_attention_layer_ids",
 ]