[bugfix] main-sd-bugfix

Signed-off-by: mengwei805 <mengwei25@huawei.com>

Author: mengwei805

vllm_ascend/attention/attention.py

@@ -113,7 +113,8 @@ def get_splitfuse_attn_mask(
             self.update_attn_cache(max_seq_len, dtype, device)
             # FIXME: Currently the mask value of chunked-prefill situation and Prefill-Only situation
             # is not the same. Fix this in the future when kernel is ready.
-            if self.attn_mask_cache[0][1] > 0:
+            if self.attn_mask_cache.numel(
+            ) > 1 and self.attn_mask_cache[0][1] > 0:
                 attn_mask = self.get_attn_mask(  # type: ignore
                     max_seq_len, dtype, device)
                 attn_mask *= -10000

vllm_ascend/attention/mla_v1.py

@@ -6,7 +6,6 @@
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionLayer,
                                               AttentionMetadata,
                                               MLAAttentionImpl)
-from vllm.logger import init_logger
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                LinearBase, RowParallelLinear,
                                                UnquantizedLinearMethod)
@@ -21,8 +20,6 @@
     from vllm.v1.core.sched.output import SchedulerOutput
     from vllm.v1.worker.gpu_input_batch import InputBatch
 
-logger = init_logger(__name__)
-
 
 class AscendMLABackend(AttentionBackend):
 

vllm_ascend/core/scheduler.py

@@ -16,14 +16,12 @@
 #
 from collections import deque
 
-from vllm.logger import init_logger
+from vllm.logger import logger
 from vllm.utils import cdiv
 from vllm.v1.core.sched.output import NewRequestData, SchedulerOutput
 from vllm.v1.core.sched.scheduler import Scheduler
 from vllm.v1.request import Request, RequestStatus
 
-logger = init_logger(__name__)
-
 
 class AscendScheduler(Scheduler):
     """This Scheduler extends vllm's original v1 scheduler

vllm_ascend/distributed/parallel_state.py

@@ -36,7 +36,6 @@ def init_ascend_model_parallel(
                                               expert_tensor_parallel_size)
 
     global _EP
-    assert _EP is None, ("expert parallel group is already initialized")
     group_ranks = []
     for i in range(num_expert_parallel_groups):
         ranks = list(range(i, world_size, num_expert_parallel_groups))
@@ -49,8 +48,6 @@ def init_ascend_model_parallel(
 
     group_ranks = []
     global _ETP
-    assert _ETP is None, (
-        "expert tensor parallel group is already initialized")
     for i in range(num_expert_tensor_parallel_groups):
         ranks = list(
             range(i * expert_tensor_parallel_size,

vllm_ascend/models/deepseek_mtp.py

@@ -1,6 +1,6 @@
 #
 # Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
-# Adapted from vllm/model_executor/models/qwen2_vl.py
+# Adapted from vllm/model_executor/models/deepseek_mtp.py
 # Copyright 2023 The vLLM team.
 #
 # This file is a part of the vllm-ascend project.
@@ -17,12 +17,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import List, Optional
+from typing import Optional
 
 import torch
 import torch.nn as nn
 from transformers import PretrainedConfig
-from vllm.attention.backends.abstract import AttentionMetadata
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
@@ -70,8 +69,6 @@ def forward(
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_cache: torch.Tensor,
-        attn_metadata: AttentionMetadata,
         previous_hidden_states: torch.Tensor,
         inputs_embeds: Optional[torch.Tensor] = None,
         spec_step_index: int = 0,
@@ -91,8 +88,6 @@ def forward(
 
         hidden_states, residual = self.mtp_block(positions=positions,
                                                  hidden_states=hidden_states,
-                                                 kv_cache=kv_cache,
-                                                 attn_metadata=attn_metadata,
                                                  residual=None)
         hidden_states = residual + hidden_states
         return hidden_states
@@ -130,8 +125,6 @@ def forward(
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
         previous_hidden_states: torch.Tensor,
         inputs_embeds: Optional[torch.Tensor] = None,
         spec_step_idx: int = 0,
@@ -140,8 +133,6 @@ def forward(
         return self.layers_list[current_step_idx](
             input_ids,
             positions,
-            kv_caches[current_step_idx],
-            attn_metadata,
             previous_hidden_states,
             inputs_embeds,
             current_step_idx,
@@ -162,6 +153,14 @@ def compute_logits(
 
 
 class CustomDeepSeekMTP(DeepSeekMTP):
+    # NOTE 1.The quantized MTP layer of deepseek on the NPU is not quantized;
+    # NOTE 2.The description file generated by the current msmodelslim tool does not have
+    # MTP layer info. Please manually add it and set the value to FLOAT.
+    packed_modules_mapping = {
+        "gate_up_proj": ["gate_proj", "up_proj"],
+        "experts":
+        ["experts.0.gate_proj", "experts.0.up_proj", "experts.0.down_proj"]
+    }
 
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         nn.Module.__init__(self)

vllm_ascend/patch/worker/patch_common/patch_spec_decode_worker.py

@@ -18,7 +18,7 @@
 from typing import Any, Dict, Optional
 
 from vllm.config import ParallelConfig
-from vllm.logger import init_logger
+from vllm.logger import logger
 from vllm.model_executor.layers.rejection_sampler import RejectionSampler
 from vllm.model_executor.layers.spec_decode_base_sampler import \
     SpecDecodeBaseSampler
@@ -34,8 +34,6 @@
 
 from vllm_ascend.worker.draft_model_runner import TP1DraftModelRunner
 
-logger = init_logger(__name__)
-
 
 def create_worker(
     cls,

vllm_ascend/quantization/quant_config.py

@@ -23,8 +23,6 @@
 from vllm.distributed import get_tensor_model_parallel_rank
 from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
                                                   FusedMoeWeightScaleSupported)
-from vllm.model_executor.layers.fused_moe.layer import \
-    UnquantizedFusedMoEMethod
 from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
                                                RowParallelLinear,
                                                UnquantizedLinearMethod)
@@ -36,6 +34,7 @@
 from vllm.model_executor.parameter import PerTensorScaleParameter
 from vllm.model_executor.utils import set_weight_attrs
 
+from ..ops.fused_moe import AscendUnquantizedFusedMoEMethod
 from .quantizer import AscendQuantizer
 
 
@@ -97,7 +96,7 @@ def get_quant_method(self, layer: torch.nn.Module,
         elif isinstance(layer, FusedMoE):
             if self.is_layer_skipped_ascend(prefix,
                                             self.packed_modules_mapping):
-                return UnquantizedFusedMoEMethod()
+                return AscendUnquantizedFusedMoEMethod()
             return AscendFusedMoEMethod(self, prefix,
                                         self.packed_modules_mapping)
         return None

vllm_ascend/worker/draft_model_runner.py

@@ -19,7 +19,7 @@
 
 import torch
 from vllm.forward_context import set_forward_context
-from vllm.logger import init_logger
+from vllm.logger import logger
 from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.multimodal import MultiModalKwargs
 from vllm.sequence import ExecuteModelRequest, IntermediateTensors
@@ -29,8 +29,6 @@
 
 from vllm_ascend.attention.attention import AscendMetadata
 
-logger = init_logger(__name__)
-
 # A flag to enable debug prints for the updated input tensors
 # before each step.
 debug_advance_input = False