[Ascend] support qwen3.5 35BA3B

lmdeploy/pytorch/backends/dlinfer/ascend/op_backend.py

@@ -381,11 +381,24 @@ def get_moe_group_name(group):
                 AscendKVQuantMeta.set_value(step_context.block_offsets.device, step_context.model_config.dtype,
                                             record_file, total_layers)
 
+        cu_seqlens = None
+        has_initial_state = None
+
+        is_gated_delta = step_context.model_config.is_gated_delta
+        if is_gated_delta:
+            q_start_loc = step_context.q_start_loc.to(dtype=step_context.q_seqlens.dtype,
+                                                      device=step_context.q_seqlens.device)
+            cu_seqlens = torch.cat((q_start_loc, step_context.q_seqlens.sum().unsqueeze(0))).int()
+            if not step_context.is_decoding:
+                has_initial_state = ~(step_context.q_seqlens == step_context.kv_seqlens)
+
         attn_meta_cls = cls.get_attention_metadata_cls()
         attn_metadata = attn_meta_cls(
             step_context.is_decoding,
             step_context.block_offsets,
-            q_start_loc=None,
+            # cu_seqlens is only used in GDN and is passed down via q_start_loc.
+            # Otherwise, q_start_loc is None.
+            q_start_loc=cu_seqlens,
             q_seqlens=q_seqlens_cpu,
             # kv_seqlens_expanded is only expanded in paged prefill,
             # otherwise it equals kv_seqlens_cpu
@@ -398,6 +411,7 @@ def get_moe_group_name(group):
             max_kv_seq_len=max_kv_seq_len,
             quant_policy=step_context.kv_quant_policy,
             quant_meta=AscendKVQuantMeta.quant_meta,
+            has_initial_state=has_initial_state,
         )
         step_context.attn_metadata = attn_metadata
 
@@ -438,7 +452,13 @@ def build_graph_runner(model: torch.nn.Module, model_config: ModelConfig, cache_
 
     @staticmethod
     def init():
-        """Initialize Ascend backend."""
+        """Initialize Ascend backend.
+
+        Note: triton device properties initialization is only required for models that use
+        linear attention (e.g. Qwen3.5 35B). If triton-ascend is not installed, a warning
+        is emitted but non-linear-attention models are unaffected.
+        """
+
         try:
             from torch_npu.contrib import transfer_to_npu  # noqa: F401
         except ImportError:
@@ -448,6 +468,16 @@ def init():
             logger.warning(f'Error during Ascend initialization: {str(e)}. '
                            'Please check your Ascend environment configuration.')
 
+        try:
+            from dlinfer.vendor.ascend.triton_ops.triton_utils import init_device_properties_triton
+            init_device_properties_triton()
+        except ImportError:
+            logger.warning('triton-ascend is not installed. Only linear attention models (e.g. Qwen3.5 35B) '
+                           'require triton-ascend. Please install it with: pip install triton-ascend==3.2.0')
+        except Exception as e:
+            logger.warning(f'Error during Ascend initialization: {str(e)}. '
+                           'Please check your Ascend environment configuration.')
+
     @staticmethod
     def ccl_backend():
         return 'hccl'

lmdeploy/pytorch/backends/dlinfer/attention.py

@@ -18,6 +18,7 @@ class DlinferAttentionMetadata(AttentionMetadata):
     max_kv_seq_len: int = 1
     quant_meta: dict = None
     cu_seq_lens_kv: Tensor | None = None
+    has_initial_state: Tensor | None = None
 
 
 class DlinferAttentionImpl(AttentionImpl[DlinferAttentionMetadata]):

lmdeploy/pytorch/config.py

@@ -451,6 +451,7 @@ def from_hf_config(
             is_draft_model=is_draft_model,
             spec_method=spec_method,
             num_spec_tokens=num_spec_tokens,
+            device_type=device_type,
         )
 
         if model_config.k_head_dim is None:

lmdeploy/pytorch/configurations/qwen3_5.py

@@ -58,7 +58,8 @@ def build(cls,
         else:
             recurrent_state_shape = (num_delta_layers, num_v_heads, head_k_dim, head_v_dim)
 
-        if is_bf16_supported():
+        device_type = kwargs.get('device_type', 'auto')
+        if is_bf16_supported(device_type):
             dtype = torch.bfloat16
         else:
             dtype = torch.float16