Upgrade to 0.11.1 newest vllm commit

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

Author: wxsIcey

vllm_ascend/kv_offload/cpu_npu.py

@@ -2,11 +2,17 @@
 import torch
 from vllm.attention import AttentionBackend
 from vllm.logger import init_logger
-from vllm.utils import is_pin_memory_available
 from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
 from vllm.v1.kv_offload.worker.worker import (OffloadingHandler,
                                               TransferResult, TransferSpec)
 
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.utils import is_pin_memory_available
+else:
+    from vllm.utils.platform_utils import is_pin_memory_available
+
 logger = init_logger(__name__)
 
 

vllm_ascend/models/qwen3_next.py

@@ -51,6 +51,8 @@
 from vllm.transformers_utils.configs import Qwen3NextConfig
 from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
 
+from vllm_ascend.utils import vllm_version_is
+
 from vllm.model_executor.models.qwen3_next import (  # isort: skip
     Qwen3NextAttention, Qwen3NextDecoderLayer, Qwen3NextForCausalLM,
     Qwen3NextGatedDeltaNet, Qwen3NextModel, Qwen3NextSparseMoeBlock,
@@ -201,7 +203,11 @@ def _forward(
         spec_query_start_loc = attn_metadata.spec_query_start_loc
         non_spec_query_start_loc = attn_metadata.non_spec_query_start_loc
         spec_sequence_masks = attn_metadata.spec_sequence_masks
-        spec_token_masks = attn_metadata.spec_token_masks
+        if vllm_version_is("0.11.0"):
+            spec_token_masks = attn_metadata.spec_token_masks
+        else:
+            spec_token_indx = attn_metadata.spec_token_indx
+            non_spec_token_indx = attn_metadata.non_spec_token_indx
         spec_state_indices_tensor = attn_metadata.spec_state_indices_tensor  # noqa: E501
         non_spec_state_indices_tensor = attn_metadata.non_spec_state_indices_tensor  # noqa: E501
         self_kv_cache = self.kv_cache[forward_context.virtual_engine]
@@ -216,8 +222,9 @@ def _forward(
 
         # 1. Set up dimensions for reshapes later
         projected_states, _ = self.in_proj(hidden_states[:num_actual_tokens])
-        if spec_token_masks is not None:
-            spec_token_masks = spec_token_masks[:num_actual_tokens]
+        if vllm_version_is("0.11.0"):
+            if spec_token_masks is not None:
+                spec_token_masks = spec_token_masks[:num_actual_tokens]
         projected_states_qkvz, projected_states_ba = torch.split(
             projected_states,
             [
@@ -242,8 +249,13 @@ def _forward(
                 mixed_qkv_spec = mixed_qkv
                 mixed_qkv_non_spec = None
             else:
-                mixed_qkv_spec = mixed_qkv[spec_token_masks]
-                mixed_qkv_non_spec = mixed_qkv[~spec_token_masks]
+                if vllm_version_is("0.11.0"):
+                    mixed_qkv_spec = mixed_qkv[spec_token_masks]
+                    mixed_qkv_non_spec = mixed_qkv[~spec_token_masks]
+                else:
+                    mixed_qkv_spec = mixed_qkv.index_select(0, spec_token_indx)
+                    mixed_qkv_non_spec = mixed_qkv.index_select(
+                        0, non_spec_token_indx)
         else:
             mixed_qkv_spec = None
             mixed_qkv_non_spec = mixed_qkv
@@ -293,10 +305,16 @@ def _forward(
                 g_non_spec = None
                 beta_non_spec = None
             else:
-                g_spec = g[:, spec_token_masks]
-                beta_spec = beta[:, spec_token_masks]
-                g_non_spec = g[:, ~spec_token_masks]
-                beta_non_spec = beta[:, ~spec_token_masks]
+                if vllm_version_is("0.11.0"):
+                    g_spec = g[:, spec_token_masks]
+                    beta_spec = beta[:, spec_token_masks]
+                    g_non_spec = g[:, ~spec_token_masks]
+                    beta_non_spec = beta[:, ~spec_token_masks]
+                else:
+                    g_spec = g.index_select(1, spec_token_indx)
+                    beta_spec = beta.index_select(1, spec_token_indx)
+                    g_non_spec = g.index_select(1, non_spec_token_indx)
+                    beta_non_spec = beta.index_select(1, non_spec_token_indx)
         else:
             g_spec = None
             beta_spec = None
@@ -404,8 +422,14 @@ def _forward(
                 dtype=core_attn_out_non_spec.dtype,
                 device=core_attn_out_non_spec.device,
             )
-            core_attn_out[:, spec_token_masks] = core_attn_out_spec
-            core_attn_out[:, ~spec_token_masks] = core_attn_out_non_spec
+            if vllm_version_is("0.11.0"):
+                core_attn_out[:, spec_token_masks] = core_attn_out_spec
+                core_attn_out[:, ~spec_token_masks] = core_attn_out_non_spec
+            else:
+                core_attn_out.index_copy_(1, spec_token_indx,
+                                          core_attn_out_spec)
+                core_attn_out.index_copy_(1, non_spec_token_indx,
+                                          core_attn_out_non_spec)
         elif spec_sequence_masks is not None:
             core_attn_out = core_attn_out_spec
         else:
@@ -673,4 +697,4 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.num_physical_experts = example_layer.n_physical_experts
         self.num_local_physical_experts = example_layer.n_local_physical_experts
         self.num_routed_experts = example_layer.n_routed_experts
-        self.num_redundant_experts = example_layer.n_redundant_experts
+        self.num_redundant_experts = example_layer.n_redundant_experts

vllm_ascend/sample/rejection_sampler.py

@@ -5,10 +5,17 @@
 import torch.nn as nn
 import vllm.v1.sample.rejection_sampler as rs
 from vllm.v1.sample.metadata import SamplingMetadata
-from vllm.v1.sample.rejection_sampler import (RejectionSampler, compute_probs,
+from vllm.v1.sample.rejection_sampler import (RejectionSampler,
                                               generate_uniform_probs)
 from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
 
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.11.0"):
+    from vllm.v1.sample.rejection_sampler import compute_probs
+else:
+    from vllm.v1.sample.rejection_sampler import apply_sampling_constraints
+
 PLACEHOLDER_TOKEN_ID = -1
 GREEDY_TEMPERATURE = -1
 # Maximum number of speculative draft tokens allowed per request in a single
@@ -82,11 +89,18 @@ def forward(
         # [num_tokens, vocab_size]
         # NOTE(woosuk): `target_logits` can be updated in place inside the
         # `compute_probs` function.
-        target_probs = compute_probs(
-            target_logits,
-            metadata.cu_num_draft_tokens,
-            sampling_metadata,
-        )
+        if vllm_version_is("0.11.0"):
+            target_probs = compute_probs(
+                target_logits,
+                metadata.cu_num_draft_tokens,
+                sampling_metadata,
+            )
+        else:
+            target_probs = apply_sampling_constraints(
+                target_logits,
+                metadata.cu_num_draft_tokens,
+                sampling_metadata,
+            )
 
         output_token_ids = rejection_sample(
             metadata.draft_token_ids,

vllm_ascend/spec_decode/eagle_proposer.py

@@ -12,7 +12,6 @@
 from vllm.model_executor.model_loader import get_model
 from vllm.model_executor.models import supports_multimodal
 from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
-from vllm.utils import is_pin_memory_available
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.sample.metadata import SamplingMetadata
 from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
@@ -27,8 +26,10 @@
 
 if vllm_version_is("0.11.0"):
     from vllm.config import CompilationLevel
+    from vllm.utils import is_pin_memory_available
 else:
     from vllm.config import CompilationMode
+    from vllm.utils.platform_utils import is_pin_memory_available
 
 PADDING_SLOT_ID = -1
 

vllm_ascend/worker/model_runner_v1.py

@@ -72,7 +72,7 @@
 from vllm.sampling_params import SamplingType
 from vllm.sequence import IntermediateTensors
 from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
-from vllm.utils import cdiv, is_pin_memory_available
+from vllm.utils import cdiv
 from vllm.utils.jsontree import json_map_leaves
 from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
 from vllm.v1.attention.backends.utils import (
@@ -155,13 +155,14 @@
 if vllm_version_is("0.11.0"):
     from vllm.attention.layer import Attention
     from vllm.config import CompilationLevel
-    from vllm.utils import LazyLoader
+    from vllm.utils import LazyLoader, is_pin_memory_available
 
     from vllm_ascend.models.layers.mla import AscendMultiHeadLatentAttention
 else:
     from vllm.attention.layer import MLAAttention
     from vllm.config import CompilationMode
     from vllm.utils.import_utils import LazyLoader
+    from vllm.utils.platform_utils import is_pin_memory_available
 
 if TYPE_CHECKING:
     import xgrammar as xgr  # type: ignore[import-untyped]