EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Drop 0.11.0 support (#4377)

Browse Source 
There is a lot hack code for v0.11.0, which makes the code hard to
upgrade to newer vLLM version. Since v0.11.0 will release soon. Let's
drop v0.11.0 support first. Then we'll upgrade to v0.11.2 soon.


- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-11-24 17:08:20 +08:00
committed by GitHub
parent 41ddb06554
commit a1f142b7ad
80 changed files with 467 additions and 1755 deletions
Download Patch File Download Diff File Expand all files Collapse all files

104
vllm_ascend/models/layers/mla.py
View File

@@ -24,32 +24,16 @@ from typing import Optional
import torch
from torch import nn
from vllm.attention import AttentionMetadata
from vllm.attention.layer import MLAAttention
from vllm.config import CacheConfig, get_current_vllm_config
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.model_executor.layers.mla import MLAModules
from vllm.model_executor.layers.mla import (MLAModules,
MultiHeadLatentAttentionWrapper)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.utils.torch_utils import direct_register_custom_op
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.utils import vllm_version_is
if vllm_version_is("0.11.0"):
from vllm.attention import Attention
from vllm.model_executor.layers.mla import \
MultiHeadLatentAttention as MultiHeadLatentAttentionWrapper
from vllm.utils import direct_register_custom_op
else:
from vllm.attention.layer import MLAAttention
from vllm.model_executor.layers.mla import MultiHeadLatentAttentionWrapper
from vllm.utils.torch_utils import direct_register_custom_op
if vllm_version_is("0.11.0"):
from vllm.attention import Attention
from vllm.model_executor.layers.mla import \
MultiHeadLatentAttention as MultiHeadLatentAttentionWrapper
else:
from vllm.attention.layer import MLAAttention
from vllm.model_executor.layers.mla import MultiHeadLatentAttentionWrapper
class IndexerWrapper(nn.Module):
@@ -81,7 +65,6 @@ class IndexerWrapper(nn.Module):
return
# TODO(whx): adapt v0.11.0 and DSA
class AscendMultiHeadLatentAttention(MultiHeadLatentAttentionWrapper):
def __init__(
@@ -119,61 +102,30 @@ class AscendMultiHeadLatentAttention(MultiHeadLatentAttentionWrapper):
ascend_indexer = IndexerWrapper(mla_modules.indexer)
else:
ascend_indexer = None
if vllm_version_is("0.11.0"):
self.mla_attn = Attention(
num_heads=num_heads,
head_size=self.kv_lora_rank + self.qk_rope_head_dim,
scale=scale,
num_kv_heads=1,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
use_mla=True,
indexer=ascend_indexer,
use_sparse=mla_modules.is_sparse,
# MLA Args
q_lora_rank=self.q_lora_rank,
kv_lora_rank=self.kv_lora_rank,
qk_nope_head_dim=self.qk_nope_head_dim,
qk_rope_head_dim=self.qk_rope_head_dim,
v_head_dim=self.v_head_dim,
qk_head_dim=self.qk_head_dim,
rotary_emb=mla_modules.rotary_emb,
fused_qkv_a_proj=mla_modules.fused_qkv_a_proj,
q_b_proj=mla_modules.q_b_proj,
q_a_layernorm=mla_modules.q_a_layernorm,
q_proj=mla_modules.q_proj,
kv_a_proj_with_mqa=mla_modules.kv_a_proj_with_mqa,
kv_a_layernorm=mla_modules.kv_a_layernorm,
kv_b_proj=mla_modules.kv_b_proj,
o_proj=mla_modules.o_proj,
)
else:
self.mla_attn = MLAAttention(
num_heads=num_heads,
scale=scale,
qk_nope_head_dim=self.qk_nope_head_dim,
qk_rope_head_dim=self.qk_rope_head_dim,
v_head_dim=self.v_head_dim,
q_lora_rank=self.q_lora_rank,
kv_lora_rank=self.kv_lora_rank,
kv_b_proj=mla_modules.kv_b_proj,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
use_sparse=mla_modules.is_sparse,
indexer=ascend_indexer,
# extra args
rotary_emb=mla_modules.rotary_emb,
fused_qkv_a_proj=mla_modules.fused_qkv_a_proj,
q_b_proj=mla_modules.q_b_proj,
q_a_layernorm=mla_modules.q_a_layernorm,
q_proj=mla_modules.q_proj,
kv_a_proj_with_mqa=mla_modules.kv_a_proj_with_mqa,
kv_a_layernorm=mla_modules.kv_a_layernorm,
o_proj=mla_modules.o_proj,
)
self.mla_attn = MLAAttention(
num_heads=num_heads,
scale=scale,
qk_nope_head_dim=self.qk_nope_head_dim,
qk_rope_head_dim=self.qk_rope_head_dim,
v_head_dim=self.v_head_dim,
q_lora_rank=self.q_lora_rank,
kv_lora_rank=self.kv_lora_rank,
kv_b_proj=mla_modules.kv_b_proj,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
use_sparse=mla_modules.is_sparse,
indexer=ascend_indexer,
# extra args
rotary_emb=mla_modules.rotary_emb,
fused_qkv_a_proj=mla_modules.fused_qkv_a_proj,
q_b_proj=mla_modules.q_b_proj,
q_a_layernorm=mla_modules.q_a_layernorm,
q_proj=mla_modules.q_proj,
kv_a_proj_with_mqa=mla_modules.kv_a_proj_with_mqa,
kv_a_layernorm=mla_modules.kv_a_layernorm,
o_proj=mla_modules.o_proj,
)
compilation_config = get_current_vllm_config().compilation_config
if prefix in compilation_config.static_forward_context:

23
vllm_ascend/models/qwen2_5_vl.py
View File

@@ -40,14 +40,11 @@ from vllm.model_executor.models.qwen2_5_vl import (
Qwen2_5_VLDummyInputsBuilder, Qwen2_5_VLForConditionalGeneration,
Qwen2_5_VLMultiModalProcessor, Qwen2_5_VLProcessingInfo)
from vllm.model_executor.models.utils import maybe_prefix
from vllm.model_executor.models.vision import conv3d_to_linear_weight
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm_ascend.ascend_forward_context import set_ascend_forward_context
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, is_enable_nz,
vllm_version_is)
if not vllm_version_is("0.11.0"):
from vllm.model_executor.models.vision import conv3d_to_linear_weight
from vllm_ascend.utils import ACL_FORMAT_FRACTAL_ND, is_enable_nz
MIN_PAD_SIZE = 64 # min_size to pad weight
MAX_PAD_SIZE = 128 # max_size to pad weight
@@ -360,9 +357,8 @@ class AscendQwen2_5_VisionTransformer(Qwen2_5_VisionTransformer):
params_dict = dict(self.named_parameters(remove_duplicate=False))
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if not vllm_version_is("0.11.0"):
if name.endswith("patch_embed.proj.weight"):
loaded_weight = conv3d_to_linear_weight(loaded_weight)
if name.endswith("patch_embed.proj.weight"):
loaded_weight = conv3d_to_linear_weight(loaded_weight)
for (param_name, weight_name, shard_id) in stacked_params_mapping:
if weight_name not in name:
continue
@@ -537,11 +533,8 @@ class AscendQwen2_5_VLForConditionalGeneration(
image_embeds = image_input["image_embeds"].type(self.visual.dtype)
else:
pixel_values = image_input["pixel_values"].type(self.visual.dtype)
if vllm_version_is("0.11.0"):
with set_ascend_forward_context(None, self.vllm_config):
image_embeds = self.visual(pixel_values, grid_thw=grid_thw)
else:
with set_ascend_forward_context(None, self.vllm_config):
image_embeds = self.visual(pixel_values, grid_thw=grid_thw)
# Split concatenated embeddings for each image item.
merge_size = self.visual.spatial_merge_size
@@ -558,13 +551,9 @@ class AscendQwen2_5_VLForConditionalGeneration(
else:
pixel_values_videos = video_input["pixel_values_videos"].type(
self.visual.dtype)
if vllm_version_is("0.11.0"):
with set_ascend_forward_context(None, self.vllm_config):
video_embeds = self.visual(pixel_values_videos,
grid_thw=grid_thw)
else:
with set_ascend_forward_context(None, self.vllm_config):
video_embeds = self.visual(pixel_values_videos,
grid_thw=grid_thw)
# Split concatenated embeddings for each video item.
merge_size = self.visual.spatial_merge_size

12
vllm_ascend/models/qwen2_vl.py
View File

@@ -38,13 +38,10 @@ from vllm.model_executor.models.qwen2_vl import (
Qwen2VLForConditionalGeneration, Qwen2VLMultiModalProcessor,
Qwen2VLProcessingInfo)
from vllm.model_executor.models.utils import maybe_prefix
from vllm.model_executor.models.vision import conv3d_to_linear_weight
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, is_enable_nz,
vllm_version_is)
if not vllm_version_is("0.11.0"):
from vllm.model_executor.models.vision import conv3d_to_linear_weight
from vllm_ascend.utils import ACL_FORMAT_FRACTAL_ND, is_enable_nz
MIN_PAD_SIZE = 64 # min_size to pad weight
MAX_PAD_SIZE = 128 # max_size to pad weight
@@ -308,9 +305,8 @@ class AscendQwen2VisionTransformer(Qwen2VisionTransformer):
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if not vllm_version_is("0.11.0"):
if name.endswith("patch_embed.proj.weight"):
loaded_weight = conv3d_to_linear_weight(loaded_weight)
if name.endswith("patch_embed.proj.weight"):
loaded_weight = conv3d_to_linear_weight(loaded_weight)
for (param_name, weight_name, shard_id) in stacked_params_mapping:
if weight_name not in name:

47
vllm_ascend/models/qwen3_next.py
View File

@@ -50,8 +50,6 @@ from vllm.model_executor.utils import set_weight_attrs
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,
@@ -202,11 +200,8 @@ class CustomQwen3NextGatedDeltaNet(Qwen3NextGatedDeltaNet, MambaBase):
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
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_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]
@@ -221,9 +216,6 @@ class CustomQwen3NextGatedDeltaNet(Qwen3NextGatedDeltaNet, MambaBase):
# 1. Set up dimensions for reshapes later
projected_states, _ = self.in_proj(hidden_states[: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,
[
@@ -248,13 +240,9 @@ class CustomQwen3NextGatedDeltaNet(Qwen3NextGatedDeltaNet, MambaBase):
mixed_qkv_spec = mixed_qkv
mixed_qkv_non_spec = None
else:
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)
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
@@ -322,16 +310,10 @@ class CustomQwen3NextGatedDeltaNet(Qwen3NextGatedDeltaNet, MambaBase):
g_non_spec = None
beta_non_spec = None
else:
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)
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
@@ -439,14 +421,9 @@ class CustomQwen3NextGatedDeltaNet(Qwen3NextGatedDeltaNet, MambaBase):
dtype=core_attn_out_non_spec.dtype,
device=core_attn_out_non_spec.device,
)
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)
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: