[Feature][Model] Add MiniMax M3 model core support (vllm-project#13009)

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

This PR adds the MiniMax M3 model core support on Ascend (split from
vllm-project#12734, based on the
implementation from
vllm-project#12448):

- Registers `MiniMaxM3SparseForCausalLM` in the Ascend model registry.
- Adds the MiniMax M3 sparse text model implementation.
- Adds MiniMax M3 MSA/indexer cache support.
- Adds a guard to reject `enable_fused_mc2=1` for MiniMax M3, which is
currently unsupported.
- Updates `model_runner_v1.py` to discover custom `AttentionLayerBase`
KV cache specs and attention backends through the existing
`get_kv_cache_spec()` and `get_attn_backend()` interfaces.

The runner change does not introduce MiniMax-specific imports into
`model_runner_v1.py`. It keeps model-specific behavior inside the
model/backend implementation and lets the runner depend on the common
`AttentionLayerBase` abstraction.

These changes provide the model-side components required to run MiniMax
M3 sparse attention on Ascend, together with the sparse-attention
operators added in
vllm-project#12833.

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

Yes. This PR introduces MiniMax M3 sparse text model support on Ascend.
It does not change an existing public CLI or API.

### How was this patch tested?

Verified MiniMax M3 service startup on 8 Ascend cards with W8A8
quantization.

Verified single curl requests:

- Text request passed.
- Image request passed.
- Video request passed.

The following checks were run:

- `bash format.sh`
- `bash format.sh ci`
- `pytest tests/ut/worker/a2/test_model_runner_v1.py`
- `pytest tests/ut/attention/a2/test_sfa_v1.py`
- `pytest tests/ut/attention/a2/test_sfa_cp_precision.py`
- `pytest tests/ut/patch/worker/test_patch_deepseek_v2.py`
- `pytest tests/ut/kv_offload/test_mooncake_connector.py -k "sfa or
indexer"`

Test results:

- `tests/ut/worker/a2/test_model_runner_v1.py`: 18 passed
- `tests/ut/attention/a2/test_sfa_v1.py`: 31 passed
- `tests/ut/attention/a2/test_sfa_cp_precision.py`: 2 passed
- `tests/ut/patch/worker/test_patch_deepseek_v2.py`: 3 passed
- `tests/ut/kv_offload/test_mooncake_connector.py -k "sfa or indexer"`:
5 passed, 99 deselected


- vLLM version: v0.25.1
- vLLM main:
vllm-project/vllm@fe784ff

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/ascend_config.py

@@ -156,6 +156,14 @@ def __init__(self, vllm_config: "VllmConfig"):
             ascend_envs.VLLM_ASCEND_ENABLE_FUSED_MC2,
         )
         assert self.enable_fused_mc2 in (0, 1), f"enable_fused_mc2 must be 0 or 1, got {self.enable_fused_mc2}"
+        model_architectures = getattr(vllm_config.model_config, "architectures", None) or []
+        assert not (
+            self.enable_fused_mc2 == 1
+            and any(architecture.startswith("MiniMaxM3") for architecture in model_architectures)
+        ), (
+            "MiniMax M3 does not support enable_fused_mc2=1. Please set "
+            "additional_config.enable_fused_mc2 to 0 or unset VLLM_ASCEND_ENABLE_FUSED_MC2."
+        )
         if self.enable_fused_mc2 == 1 and self.multistream_overlap_shared_expert:
             self.multistream_overlap_shared_expert = False
             logger.warning_once(

vllm_ascend/models/__init__.py

@@ -3,6 +3,10 @@
 
 def register_model():
     ModelRegistry.register_model("DeepseekV4ForCausalLM", "vllm_ascend.models.deepseek_v4:AscendDeepseekV4ForCausalLM")
+    ModelRegistry.register_model(
+        "MiniMaxM3SparseForCausalLM",
+        "vllm_ascend.models.minimax_m3:MiniMaxM3SparseForCausalLM",
+    )
     ModelRegistry.register_model("DeepSeekV4MTPModel", "vllm_ascend.models.deepseek_v4_mtp:DeepSeekV4MTP")
     ModelRegistry.register_model(
         "DSparkDraftModel",

vllm_ascend/models/minimax_m3/__init__.py

@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm_ascend.models.minimax_m3.minimax_m3 import (
+    MiniMaxM3Attention,
+    MiniMaxM3MoE,
+    MiniMaxM3SparseAttention,
+    MiniMaxM3SparseForCausalLM,
+    _get_rope_parameters,
+    _sparse_attention_layer_ids,
+)
+
+__all__ = [
+    "MiniMaxM3Attention",
+    "MiniMaxM3MoE",
+    "MiniMaxM3SparseAttention",
+    "MiniMaxM3SparseForCausalLM",
+    "_get_rope_parameters",
+    "_sparse_attention_layer_ids",
+]