[5/N][Attention] Finish eliminating vllm/attention folder (#32064)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2026-01-27 10:02:51 -05:00
committed by GitHub
parent 1f3a2c2944
commit a608b4c6c2
151 changed files with 585 additions and 527 deletions
+2 -1
View File
@@ -640,8 +640,9 @@ steps:
# grade: Blocking
source_file_dependencies:
- csrc/attention/
- vllm/attention
- vllm/v1/attention
# TODO: remove this dependency (https://github.com/vllm-project/vllm/issues/32267)
- vllm/model_executor/layers/attention
- tests/kernels/attention
commands:
- pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+2 -1
View File
@@ -568,8 +568,9 @@ steps:
mirror_hardwares: [amdexperimental]
source_file_dependencies:
- csrc/attention/
- vllm/attention
- vllm/v1/attention
# TODO: remove this dependency (https://github.com/vllm-project/vllm/issues/32267)
- vllm/model_executor/layers/attention
- tests/kernels/attention
commands:
- pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+2 -1
View File
@@ -15,8 +15,9 @@ steps:
timeout_in_minutes: 35
source_file_dependencies:
- csrc/attention/
- vllm/attention
- vllm/v1/attention
# TODO: remove this dependency (https://github.com/vllm-project/vllm/issues/32267)
- vllm/model_executor/layers/attention
- tests/kernels/attention
commands:
- pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+1 -1
View File
@@ -2,8 +2,8 @@
# for more info about CODEOWNERS file
# This lists cover the "core" components of vLLM that require careful review
/vllm/attention @LucasWilkinson
/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @njhill @22quinn
/vllm/model_executor/layers/attention @LucasWilkinson
/vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
/vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
/vllm/model_executor/layers/mamba @tdoublep
+1 -1
View File
@@ -29,7 +29,7 @@ The initialization code should look like this:
```python
from torch import nn
from vllm.config import VllmConfig
from vllm.attention.layer import Attention
from vllm.model_executor.layers.attention import Attention
class MyAttention(nn.Module):
def __init__(self, vllm_config: VllmConfig, prefix: str):
+1 -1
View File
@@ -271,7 +271,7 @@ Taking `MMEncoderAttention` as an example:
??? code
```python
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.custom_op import CustomOp
+1 -1
View File
@@ -21,7 +21,6 @@ from tests.compile.fusion_test_utils import (
from tests.utils import flat_product
from tests.v1.attention.utils import BatchSpec, create_common_attn_metadata
from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
from vllm.attention.layer import Attention
from vllm.compilation.fusion_attn import ATTN_OP, AttnFusionPass
from vllm.compilation.fx_utils import find_op_nodes
from vllm.compilation.matcher_utils import QUANT_OPS
@@ -40,6 +39,7 @@ from vllm.config import (
set_current_vllm_config,
)
from vllm.forward_context import get_forward_context, set_forward_context
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.quantization.utils.quant_utils import (
QuantKey,
kFp8StaticTensorSym,
+1 -1
View File
@@ -5,7 +5,6 @@ import pytest
import torch
from tests.compile.backend import TestBackend
from vllm.attention.layer import Attention
from vllm.compilation.matcher_utils import FLASHINFER_ROTARY_OP, RMS_OP, ROTARY_OP
from vllm.compilation.noop_elimination import NoOpEliminationPass
from vllm.compilation.post_cleanup import PostCleanupPass
@@ -21,6 +20,7 @@ from vllm.config import (
VllmConfig,
set_current_vllm_config,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
from vllm.platforms import current_platform
+1 -2
View File
@@ -9,8 +9,7 @@ import torch
from tests.kernels.allclose_default import get_default_atol, get_default_rtol
from tests.kernels.utils import opcheck
from vllm import _custom_ops as ops
from vllm.attention.layer import Attention
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import Attention, MMEncoderAttention
from vllm.platforms import current_platform
from vllm.utils.mem_utils import get_max_shared_memory_bytes
from vllm.utils.torch_utils import set_random_seed
+1 -1
View File
@@ -12,7 +12,7 @@ from unittest.mock import patch
import pytest
import torch
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.platforms import current_platform
from vllm.platforms.cpu import CpuPlatform
from vllm.platforms.cuda import CudaPlatform
+1 -1
View File
@@ -5,7 +5,6 @@ import numpy as np
import pytest
import torch
from vllm.attention.layer import Attention
from vllm.config import (
AttentionConfig,
CacheConfig,
@@ -19,6 +18,7 @@ from vllm.distributed.parallel_state import (
init_distributed_environment,
initialize_model_parallel,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
from vllm.platforms import current_platform
from vllm.sampling_params import SamplingParams
+3 -3
View File
@@ -7,7 +7,7 @@ from vllm.v1.worker.utils import bind_kv_cache
def test_bind_kv_cache(default_vllm_config):
from vllm.attention.layer import Attention
from vllm.model_executor.layers.attention import Attention
ctx = {
"layers.0.self_attn": Attention(32, 128, 0.1, prefix="layers.0.self_attn"),
@@ -35,7 +35,7 @@ def test_bind_kv_cache(default_vllm_config):
def test_bind_kv_cache_non_attention(default_vllm_config):
from vllm.attention.layer import Attention
from vllm.model_executor.layers.attention import Attention
# example from Jamba PP=2
ctx = {
@@ -58,7 +58,7 @@ def test_bind_kv_cache_non_attention(default_vllm_config):
def test_bind_kv_cache_draft_model(default_vllm_config):
from vllm.attention.layer import Attention
from vllm.model_executor.layers.attention import Attention
layer_names = [
"model.layers.0.attn",
-1
View File
@@ -58,7 +58,6 @@ FILES = [
SEPARATE_GROUPS = [
"tests",
# v0 related
"vllm/attention",
"vllm/compilation",
"vllm/lora",
"vllm/model_executor",
View File
View File
-33
View File
@@ -1,33 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
def validate_kv_sharing_target(
current_layer_name, target_layer_name, static_forward_context
):
error_msg = (
f"Specified KV sharing target layer for {current_layer_name} "
f"is not valid: target layer {target_layer_name} "
)
if current_layer_name == target_layer_name:
raise ValueError(error_msg + "cannot be the same as the current layer.")
if target_layer_name not in static_forward_context:
from vllm.model_executor.models.utils import extract_layer_index
# If target layer name is not in the static fwd context, it means either
# a) the target layer does not come BEFORE the current layer, or
# b) the target layer is not an Attention layer that exists in the model
current_layer_idx = extract_layer_index(current_layer_name)
target_layer_idx = extract_layer_index(target_layer_name)
if current_layer_idx <= target_layer_idx:
raise ValueError(error_msg + "must come before the current layer.")
else:
raise ValueError(error_msg + "is not a valid Attention layer in the model.")
# Currently KV sharing is only supported between layers of the same type
target_layer_attn_type = static_forward_context[target_layer_name].attn_type
expected = static_forward_context[current_layer_name].attn_type
if target_layer_attn_type != expected:
raise ValueError(
error_msg + f"must be the same type as the current layer ({expected})."
)
+1 -1
View File
@@ -11,9 +11,9 @@ from torch import fx
from torch._higher_order_ops.auto_functionalize import auto_functionalized
from torch._inductor.pattern_matcher import PatternMatcherPass
from vllm.attention.layer import Attention
from vllm.config import VllmConfig, get_layers_from_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.quantization.utils.quant_utils import (
QuantKey,
kNvfp4Dynamic,
+1 -1
View File
@@ -10,9 +10,9 @@ from torch import fx
from torch._higher_order_ops.auto_functionalize import auto_functionalized
from torch._inductor.pattern_matcher import PatternMatcherPass
from vllm.attention.layer import Attention
from vllm.config import VllmConfig, get_layers_from_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
from .fusion import empty_bf16, empty_fp32, empty_i64
@@ -8,7 +8,6 @@ from typing import Any
import torch
from vllm.attention.layer import Attention
from vllm.config import VllmConfig, get_layers_from_vllm_config
from vllm.distributed.kv_events import BlockRemoved, BlockStored, KVCacheEvent
from vllm.distributed.kv_transfer.kv_connector.utils import yield_req_data
@@ -25,6 +24,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.metrics import (
)
from vllm.forward_context import ForwardContext
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.v1.attention.backend import AttentionBackend, AttentionMetadata
from vllm.v1.core.kv_cache_manager import KVCacheBlocks
from vllm.v1.core.kv_cache_utils import BlockHash
@@ -0,0 +1,26 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.model_executor.layers.attention.attention import Attention
from vllm.model_executor.layers.attention.chunked_local_attention import (
ChunkedLocalAttention,
)
from vllm.model_executor.layers.attention.cross_attention import CrossAttention
from vllm.model_executor.layers.attention.encoder_only_attention import (
EncoderOnlyAttention,
)
from vllm.model_executor.layers.attention.mla_attention import MLAAttention
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention.static_sink_attention import (
StaticSinkAttention,
)
__all__ = [
"Attention",
"ChunkedLocalAttention",
"CrossAttention",
"EncoderOnlyAttention",
"MLAAttention",
"MMEncoderAttention",
"StaticSinkAttention",
]
@@ -1,23 +1,22 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Attention layer."""
from typing import cast
from typing import TYPE_CHECKING
import torch
import torch.nn as nn
import vllm.envs as envs
from vllm.attention.utils.kv_sharing_utils import validate_kv_sharing_target
from vllm.attention.utils.kv_transfer_utils import maybe_transfer_kv_layer
from vllm.config import CacheConfig, get_current_vllm_config
from vllm.config.vllm import VllmConfig
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.layers.attention.kv_transfer_utils import (
maybe_transfer_kv_layer,
)
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.batch_invariant import vllm_is_batch_invariant
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
UnquantizedLinearMethod,
)
from vllm.model_executor.layers.quantization import QuantizationConfig
@@ -33,20 +32,54 @@ from vllm.utils.torch_utils import (
from vllm.v1.attention.backend import (
AttentionBackend,
AttentionType,
MLAAttentionImpl,
)
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.selector import get_attn_backend
from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
KVCacheSpec,
MLAAttentionSpec,
SlidingWindowSpec,
)
if TYPE_CHECKING:
from vllm.model_executor.layers.attention import MLAAttention
logger = init_logger(__name__)
def validate_kv_sharing_target(
current_layer_name, target_layer_name, static_forward_context
):
error_msg = (
f"Specified KV sharing target layer for {current_layer_name} "
f"is not valid: target layer {target_layer_name} "
)
if current_layer_name == target_layer_name:
raise ValueError(error_msg + "cannot be the same as the current layer.")
if target_layer_name not in static_forward_context:
from vllm.model_executor.models.utils import extract_layer_index
# If target layer name is not in the static fwd context, it means either
# a) the target layer does not come BEFORE the current layer, or
# b) the target layer is not an Attention layer that exists in the model
current_layer_idx = extract_layer_index(current_layer_name)
target_layer_idx = extract_layer_index(target_layer_name)
if current_layer_idx <= target_layer_idx:
raise ValueError(error_msg + "must come before the current layer.")
else:
raise ValueError(error_msg + "is not a valid Attention layer in the model.")
# Currently KV sharing is only supported between layers of the same type
target_layer_attn_type = static_forward_context[target_layer_name].attn_type
expected = static_forward_context[current_layer_name].attn_type
if target_layer_attn_type != expected:
raise ValueError(
error_msg + f"must be the same type as the current layer ({expected})."
)
def should_load_quant_weights(quant_method: QuantizeMethodBase | None) -> bool:
"""Returns whether the quantization method should load quantized weights."""
return quant_method is not None and not isinstance(
@@ -493,236 +526,6 @@ class Attention(nn.Module, AttentionLayerBase):
)
class MLAAttention(nn.Module, AttentionLayerBase):
"""Multi-Head Latent Attention layer.
This class takes query, and compressed key/value tensors as input.
The class does the following:
1. Store the input key and value tensors in the KV cache.
2. Perform (multi-head/multi-query/grouped-query) attention.
3. Return the output tensor.
"""
def __init__(
self,
num_heads: int,
scale: float,
qk_nope_head_dim: int,
qk_rope_head_dim: int,
v_head_dim: int,
q_lora_rank: int | None,
kv_lora_rank: int,
kv_b_proj: ColumnParallelLinear,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
use_sparse: bool = False,
indexer: object | None = None,
**extra_impl_args,
):
super().__init__()
self.num_heads = num_heads
self.scale = scale
self.qk_nope_head_dim = qk_nope_head_dim
self.qk_rope_head_dim = qk_rope_head_dim
self.v_head_dim = v_head_dim
self.q_lora_rank = q_lora_rank
self.kv_lora_rank = kv_lora_rank
self.head_size = kv_lora_rank + qk_rope_head_dim
self.layer_name = prefix
if cache_config is not None:
kv_cache_dtype = cache_config.cache_dtype
block_size = cache_config.block_size
calculate_kv_scales = cache_config.calculate_kv_scales
else:
kv_cache_dtype = "auto"
block_size = 16
calculate_kv_scales = False
self.quant_config = quant_config
# Initialize KV cache quantization attributes
self.kv_cache_dtype = kv_cache_dtype
self.calculate_kv_scales = calculate_kv_scales
_init_kv_cache_quant(self, quant_config, prefix)
dtype = torch.get_default_dtype()
self.attn_backend = get_attn_backend(
self.head_size,
dtype,
kv_cache_dtype,
block_size,
use_mla=True,
use_sparse=use_sparse,
)
if (
cache_config is not None
and cache_config.enable_prefix_caching
and vllm_is_batch_invariant()
and (
self.attn_backend.get_name() == "TRITON_MLA"
or self.attn_backend.get_name() == "FLASHINFER"
)
):
logger.warning_once(
"Disabling prefix caching for TRITON_MLA / FLASHINFER "
"with batch invariance, as it is not yet supported.",
scope="local",
)
cache_config.enable_prefix_caching = False
impl_cls = cast(type[MLAAttentionImpl], self.attn_backend.get_impl_cls())
self.impl = impl_cls(
num_heads=self.num_heads,
head_size=self.head_size,
scale=self.scale,
num_kv_heads=1,
alibi_slopes=None,
sliding_window=None,
kv_cache_dtype=self.kv_cache_dtype,
logits_soft_cap=None,
attn_type=AttentionType.DECODER,
kv_sharing_target_layer_name=None,
# 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,
qk_head_dim=self.qk_nope_head_dim + self.qk_rope_head_dim,
v_head_dim=self.v_head_dim,
kv_b_proj=kv_b_proj,
indexer=indexer,
**extra_impl_args,
)
self.use_direct_call = not current_platform.opaque_attention_op()
compilation_config = get_current_vllm_config().compilation_config
if prefix in compilation_config.static_forward_context:
raise ValueError(f"Duplicate layer name: {prefix}")
compilation_config.static_forward_context[prefix] = self
self.kv_cache = [
torch.tensor([])
for _ in range(
get_current_vllm_config().parallel_config.pipeline_parallel_size
)
]
self.use_sparse = use_sparse
# Initialize q/k/v range constants.
self.q_range = torch.tensor(envs.Q_SCALE_CONSTANT, dtype=torch.float32)
self.k_range = torch.tensor(envs.K_SCALE_CONSTANT, dtype=torch.float32)
self.v_range = torch.tensor(envs.V_SCALE_CONSTANT, dtype=torch.float32)
def forward(
self,
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
output_shape: torch.Size | None = None,
) -> torch.Tensor:
if self.calculate_kv_scales:
torch.ops.vllm.maybe_calc_kv_scales(q, kv_c_normed, k_pe, self.layer_name)
if self.use_direct_call:
forward_context: ForwardContext = get_forward_context()
attn_metadata = forward_context.attn_metadata
if isinstance(attn_metadata, dict):
attn_metadata = attn_metadata[self.layer_name]
self_kv_cache = self.kv_cache[forward_context.virtual_engine]
if self.attn_backend.accept_output_buffer:
output = torch.empty(output_shape, dtype=q.dtype, device=q.device)
self.impl.forward(
self,
q,
kv_c_normed,
k_pe,
self_kv_cache,
attn_metadata,
output=output,
)
return output
else:
return self.impl.forward(
self, q, kv_c_normed, k_pe, self_kv_cache, attn_metadata
)
else:
if self.attn_backend.accept_output_buffer:
output = torch.empty(output_shape, dtype=q.dtype, device=q.device)
torch.ops.vllm.unified_mla_attention_with_output(
q,
kv_c_normed,
k_pe,
output,
self.layer_name,
)
return output
else:
return torch.ops.vllm.unified_mla_attention(
q,
kv_c_normed,
k_pe,
self.layer_name,
)
def process_weights_after_loading(self, act_dtype: torch.dtype):
if hasattr(self.impl, "process_weights_after_loading"):
self.impl.process_weights_after_loading(act_dtype)
# If we should not load quant weights, we initialize the scales to 1.0
# as the default value. See [Note: Register q/k/v/prob scales in state dict]
# for more details.
quant_method = (
self.quant_config.get_quant_method(self, prefix=self.layer_name)
if self.quant_config
else None
)
if not should_load_quant_weights(quant_method):
set_default_quant_scales(self, register_buffer=False)
def calc_kv_scales(
self, q: torch.Tensor, kv_c_normed: torch.Tensor, k_pe: torch.Tensor
) -> None:
"""Optional scale calculation for MLA inputs.
Mirrors Attention.calc_kv_scales. Not all MLA backends require this
"""
# Use safe defaults if ranges are not present
q_range = getattr(self, "q_range", torch.tensor(1.0))
k_range = getattr(self, "k_range", torch.tensor(1.0))
v_range = getattr(self, "v_range", torch.tensor(1.0))
self._q_scale.copy_(torch.abs(q).max() / q_range)
# kv_c_normed is the compressed KV representation; use it for k/v
kv_abs_max = torch.abs(kv_c_normed).max()
self._k_scale.copy_(kv_abs_max / k_range)
self._v_scale.copy_(kv_abs_max / v_range)
self._q_scale_float = self._q_scale.item()
self._k_scale_float = self._k_scale.item()
self._v_scale_float = self._v_scale.item()
self.calculate_kv_scales = False
def get_attn_backend(self) -> type[AttentionBackend]:
return self.attn_backend
def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
kv_cache_dtype = kv_cache_dtype_str_to_dtype(
self.kv_cache_dtype, vllm_config.model_config
)
return MLAAttentionSpec(
block_size=vllm_config.cache_config.block_size,
num_kv_heads=1,
head_size=self.head_size,
dtype=kv_cache_dtype,
cache_dtype_str=vllm_config.cache_config.cache_dtype,
)
def maybe_calc_kv_scales(
query: torch.Tensor,
key: torch.Tensor,
@@ -759,7 +562,7 @@ direct_register_custom_op(
def get_attention_context(
layer_name: str,
) -> tuple[dict | object | None, Attention | MLAAttention, torch.Tensor]:
) -> tuple[dict | object | None, "Attention | MLAAttention", torch.Tensor]:
"""Extract attention context for a given layer.
This helper function extracts the attention metadata, attention layer
@@ -782,7 +585,7 @@ def get_attention_context(
attn_metadata = forward_context.attn_metadata
if isinstance(attn_metadata, dict):
attn_metadata = attn_metadata[layer_name]
attn_layer: Attention | MLAAttention = forward_context.no_compile_layers[layer_name]
attn_layer = forward_context.no_compile_layers[layer_name]
kv_cache = attn_layer.kv_cache[forward_context.virtual_engine]
return attn_metadata, attn_layer, kv_cache
@@ -914,79 +717,3 @@ direct_register_custom_op(
mutates_args=["output", "output_block_scale"],
fake_impl=unified_attention_with_output_fake,
)
@maybe_transfer_kv_layer
def unified_mla_attention(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
layer_name: str,
) -> torch.Tensor:
attn_metadata, self, kv_cache = get_attention_context(layer_name)
output = self.impl.forward(self, q, kv_c_normed, k_pe, kv_cache, attn_metadata)
return output
def unified_mla_attention_fake(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
layer_name: str,
) -> torch.Tensor:
return torch.empty_like(q).contiguous()
direct_register_custom_op(
op_name="unified_mla_attention",
op_func=unified_mla_attention,
mutates_args=[],
fake_impl=unified_mla_attention_fake,
dispatch_key=current_platform.dispatch_key,
)
@maybe_transfer_kv_layer
def unified_mla_attention_with_output(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
output: torch.Tensor,
layer_name: str,
output_scale: torch.Tensor | None = None,
output_block_scale: torch.Tensor | None = None,
) -> None:
attn_metadata, self, kv_cache = get_attention_context(layer_name)
self.impl.forward(
self,
q,
kv_c_normed,
k_pe,
kv_cache,
attn_metadata,
output=output,
output_scale=output_scale,
output_block_scale=output_block_scale,
)
def unified_mla_attention_with_output_fake(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
output: torch.Tensor,
layer_name: str,
output_scale: torch.Tensor | None = None,
output_block_scale: torch.Tensor | None = None,
) -> None:
return
direct_register_custom_op(
op_name="unified_mla_attention_with_output",
op_func=unified_mla_attention_with_output,
mutates_args=["output", "output_block_scale"],
fake_impl=unified_mla_attention_with_output_fake,
dispatch_key=current_platform.dispatch_key,
)
@@ -4,9 +4,9 @@ import functools
import torch
from vllm.attention.layer import Attention
from vllm.config import CacheConfig
from vllm.config.vllm import VllmConfig
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.v1.attention.backend import (
AttentionBackend,
@@ -6,9 +6,9 @@ from copy import copy
import numpy as np
import torch
from vllm.attention.layer import Attention
from vllm.config import CacheConfig, VllmConfig
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.utils.math_utils import cdiv
from vllm.v1.attention.backend import (
AttentionBackend,
@@ -5,9 +5,9 @@ from copy import copy
import torch
from vllm.attention.layer import Attention
from vllm.config import CacheConfig
from vllm.config.vllm import VllmConfig
from vllm.model_executor.layers.attention import Attention
from vllm.v1.attention.backend import (
AttentionBackend,
AttentionMetadata,
@@ -19,7 +19,7 @@ def maybe_transfer_kv_layer(func: Callable) -> Callable:
On exit: saves the KV layer to the connector.
"""
# Import at runtime to avoid circular dependency
from vllm.attention.layer import get_attention_context
from vllm.model_executor.layers.attention.attention import get_attention_context
# Inspect the signature ONCE when the decorator is applied.
sig = inspect.signature(func)
+354 -17
View File
@@ -191,24 +191,38 @@ import functools
from abc import abstractmethod
from dataclasses import dataclass, field
from enum import Enum
from typing import ClassVar, Generic, TypeVar
from typing import TYPE_CHECKING, ClassVar, Generic, TypeVar, cast
if TYPE_CHECKING:
from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
import torch
import torch.nn as nn
from tqdm import tqdm
import vllm.envs as envs
from vllm import _custom_ops as ops
from vllm import envs
from vllm._aiter_ops import rocm_aiter_ops
from vllm.config import ModelConfig, VllmConfig, get_current_vllm_config
from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config
from vllm.distributed.parallel_state import get_dcp_group, is_global_first_rank
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.custom_op import CustomOp
from vllm.model_executor.layers.batch_invariant import (
vllm_is_batch_invariant,
from vllm.model_executor.layers.attention.attention import (
_init_kv_cache_quant,
get_attention_context,
set_default_quant_scales,
should_load_quant_weights,
)
from vllm.model_executor.layers.attention.kv_transfer_utils import (
maybe_transfer_kv_layer,
)
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.batch_invariant import vllm_is_batch_invariant
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
from vllm.model_executor.layers.quantization.utils.quant_utils import (
GroupShape,
@@ -217,11 +231,16 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_nvidia_artifactory
from vllm.utils.math_utils import cdiv, round_down
from vllm.utils.torch_utils import (
direct_register_custom_op,
kv_cache_dtype_str_to_dtype,
)
from vllm.v1.attention.backend import (
AttentionBackend,
AttentionLayer,
AttentionMetadata,
AttentionMetadataBuilder,
AttentionType,
CommonAttentionMetadata,
MLAAttentionImpl,
)
@@ -234,7 +253,320 @@ from vllm.v1.attention.backends.utils import (
)
from vllm.v1.attention.ops.common import cp_lse_ag_out_rs
from vllm.v1.attention.ops.merge_attn_states import merge_attn_states
from vllm.v1.kv_cache_interface import AttentionSpec
from vllm.v1.attention.selector import get_attn_backend
from vllm.v1.kv_cache_interface import (
AttentionSpec,
KVCacheSpec,
MLAAttentionSpec,
)
logger = init_logger(__name__)
class MLAAttention(nn.Module, AttentionLayerBase):
"""Multi-Head Latent Attention layer.
This class takes query, and compressed key/value tensors as input.
The class does the following:
1. Store the input key and value tensors in the KV cache.
2. Perform (multi-head/multi-query/grouped-query) attention.
3. Return the output tensor.
"""
def __init__(
self,
num_heads: int,
scale: float,
qk_nope_head_dim: int,
qk_rope_head_dim: int,
v_head_dim: int,
q_lora_rank: int | None,
kv_lora_rank: int,
kv_b_proj: ColumnParallelLinear,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
use_sparse: bool = False,
indexer: object | None = None,
**extra_impl_args,
):
super().__init__()
self.num_heads = num_heads
self.scale = scale
self.qk_nope_head_dim = qk_nope_head_dim
self.qk_rope_head_dim = qk_rope_head_dim
self.v_head_dim = v_head_dim
self.q_lora_rank = q_lora_rank
self.kv_lora_rank = kv_lora_rank
self.head_size = kv_lora_rank + qk_rope_head_dim
self.layer_name = prefix
if cache_config is not None:
kv_cache_dtype = cache_config.cache_dtype
block_size = cache_config.block_size
calculate_kv_scales = cache_config.calculate_kv_scales
else:
kv_cache_dtype = "auto"
block_size = 16
calculate_kv_scales = False
self.quant_config = quant_config
# Initialize KV cache quantization attributes
self.kv_cache_dtype = kv_cache_dtype
self.calculate_kv_scales = calculate_kv_scales
_init_kv_cache_quant(self, quant_config, prefix)
dtype = torch.get_default_dtype()
self.attn_backend = get_attn_backend(
self.head_size,
dtype,
kv_cache_dtype,
block_size,
use_mla=True,
use_sparse=use_sparse,
)
if (
cache_config is not None
and cache_config.enable_prefix_caching
and vllm_is_batch_invariant()
and (
self.attn_backend.get_name() == "TRITON_MLA"
or self.attn_backend.get_name() == "FLASHINFER"
)
):
logger.warning_once(
"Disabling prefix caching for TRITON_MLA / FLASHINFER "
"with batch invariance, as it is not yet supported.",
scope="local",
)
cache_config.enable_prefix_caching = False
impl_cls = cast(type[MLAAttentionImpl], self.attn_backend.get_impl_cls())
self.impl = impl_cls(
num_heads=self.num_heads,
head_size=self.head_size,
scale=self.scale,
num_kv_heads=1,
alibi_slopes=None,
sliding_window=None,
kv_cache_dtype=self.kv_cache_dtype,
logits_soft_cap=None,
attn_type=AttentionType.DECODER,
kv_sharing_target_layer_name=None,
# 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,
qk_head_dim=self.qk_nope_head_dim + self.qk_rope_head_dim,
v_head_dim=self.v_head_dim,
kv_b_proj=kv_b_proj,
indexer=indexer,
**extra_impl_args,
)
self.use_direct_call = not current_platform.opaque_attention_op()
compilation_config = get_current_vllm_config().compilation_config
if prefix in compilation_config.static_forward_context:
raise ValueError(f"Duplicate layer name: {prefix}")
compilation_config.static_forward_context[prefix] = self
self.kv_cache = [
torch.tensor([])
for _ in range(
get_current_vllm_config().parallel_config.pipeline_parallel_size
)
]
self.use_sparse = use_sparse
# Initialize q/k/v range constants.
self.q_range = torch.tensor(envs.Q_SCALE_CONSTANT, dtype=torch.float32)
self.k_range = torch.tensor(envs.K_SCALE_CONSTANT, dtype=torch.float32)
self.v_range = torch.tensor(envs.V_SCALE_CONSTANT, dtype=torch.float32)
def forward(
self,
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
output_shape: torch.Size | None = None,
) -> torch.Tensor:
if self.calculate_kv_scales:
torch.ops.vllm.maybe_calc_kv_scales(q, kv_c_normed, k_pe, self.layer_name)
if self.use_direct_call:
forward_context: ForwardContext = get_forward_context()
attn_metadata = forward_context.attn_metadata
if isinstance(attn_metadata, dict):
attn_metadata = attn_metadata[self.layer_name]
self_kv_cache = self.kv_cache[forward_context.virtual_engine]
if self.attn_backend.accept_output_buffer:
output = torch.empty(output_shape, dtype=q.dtype, device=q.device)
self.impl.forward(
self,
q,
kv_c_normed,
k_pe,
self_kv_cache,
attn_metadata,
output=output,
)
return output
else:
return self.impl.forward(
self, q, kv_c_normed, k_pe, self_kv_cache, attn_metadata
)
else:
if self.attn_backend.accept_output_buffer:
output = torch.empty(output_shape, dtype=q.dtype, device=q.device)
torch.ops.vllm.unified_mla_attention_with_output(
q,
kv_c_normed,
k_pe,
output,
self.layer_name,
)
return output
else:
return torch.ops.vllm.unified_mla_attention(
q,
kv_c_normed,
k_pe,
self.layer_name,
)
def process_weights_after_loading(self, act_dtype: torch.dtype):
if hasattr(self.impl, "process_weights_after_loading"):
self.impl.process_weights_after_loading(act_dtype)
# If we should not load quant weights, we initialize the scales to 1.0
# as the default value. See [Note: Register q/k/v/prob scales in state dict]
# for more details.
quant_method = (
self.quant_config.get_quant_method(self, prefix=self.layer_name)
if self.quant_config
else None
)
if not should_load_quant_weights(quant_method):
set_default_quant_scales(self, register_buffer=False)
def calc_kv_scales(
self, q: torch.Tensor, kv_c_normed: torch.Tensor, k_pe: torch.Tensor
) -> None:
"""Optional scale calculation for MLA inputs.
Mirrors Attention.calc_kv_scales. Not all MLA backends require this
"""
# Use safe defaults if ranges are not present
q_range = getattr(self, "q_range", torch.tensor(1.0))
k_range = getattr(self, "k_range", torch.tensor(1.0))
v_range = getattr(self, "v_range", torch.tensor(1.0))
self._q_scale.copy_(torch.abs(q).max() / q_range)
# kv_c_normed is the compressed KV representation; use it for k/v
kv_abs_max = torch.abs(kv_c_normed).max()
self._k_scale.copy_(kv_abs_max / k_range)
self._v_scale.copy_(kv_abs_max / v_range)
self._q_scale_float = self._q_scale.item()
self._k_scale_float = self._k_scale.item()
self._v_scale_float = self._v_scale.item()
self.calculate_kv_scales = False
def get_attn_backend(self) -> type[AttentionBackend]:
return self.attn_backend
def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
kv_cache_dtype = kv_cache_dtype_str_to_dtype(
self.kv_cache_dtype, vllm_config.model_config
)
return MLAAttentionSpec(
block_size=vllm_config.cache_config.block_size,
num_kv_heads=1,
head_size=self.head_size,
dtype=kv_cache_dtype,
cache_dtype_str=vllm_config.cache_config.cache_dtype,
)
@maybe_transfer_kv_layer
def unified_mla_attention(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
layer_name: str,
) -> torch.Tensor:
attn_metadata, self, kv_cache = get_attention_context(layer_name)
output = self.impl.forward(self, q, kv_c_normed, k_pe, kv_cache, attn_metadata)
return output
def unified_mla_attention_fake(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
layer_name: str,
) -> torch.Tensor:
return torch.empty_like(q).contiguous()
direct_register_custom_op(
op_name="unified_mla_attention",
op_func=unified_mla_attention,
mutates_args=[],
fake_impl=unified_mla_attention_fake,
dispatch_key=current_platform.dispatch_key,
)
@maybe_transfer_kv_layer
def unified_mla_attention_with_output(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
output: torch.Tensor,
layer_name: str,
output_scale: torch.Tensor | None = None,
output_block_scale: torch.Tensor | None = None,
) -> None:
attn_metadata, self, kv_cache = get_attention_context(layer_name)
self.impl.forward(
self,
q,
kv_c_normed,
k_pe,
kv_cache,
attn_metadata,
output=output,
output_scale=output_scale,
output_block_scale=output_block_scale,
)
def unified_mla_attention_with_output_fake(
q: torch.Tensor,
kv_c_normed: torch.Tensor,
k_pe: torch.Tensor,
output: torch.Tensor,
layer_name: str,
output_scale: torch.Tensor | None = None,
output_block_scale: torch.Tensor | None = None,
) -> None:
return
direct_register_custom_op(
op_name="unified_mla_attention_with_output",
op_func=unified_mla_attention_with_output,
mutates_args=["output", "output_block_scale"],
fake_impl=unified_mla_attention_with_output_fake,
dispatch_key=current_platform.dispatch_key,
)
class QueryLenSupport(Enum):
@@ -266,15 +598,12 @@ except ImportError:
from flash_attn import flash_attn_varlen_func # type: ignore[no-redef]
is_vllm_fa = False
try:
from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
from flashinfer.prefill import cudnn_batch_prefill_with_kv_cache # noqa: F401
flashinfer_available = True
except ImportError:
BatchPrefillWithRaggedKVCacheWrapper = object
@functools.cache
def flashinfer_available() -> bool:
import importlib.util
flashinfer_available = False
return importlib.util.find_spec("flashinfer") is not None
def dynamic_per_batched_tensor_quant(
@@ -398,8 +727,8 @@ class MLACommonPrefillMetadata:
@dataclass
class FlashInferPrefillMetadata(MLACommonPrefillMetadata):
prefill_main: BatchPrefillWithRaggedKVCacheWrapper | None = None
prefill_chunks: list[BatchPrefillWithRaggedKVCacheWrapper] = field(
prefill_main: "BatchPrefillWithRaggedKVCacheWrapper | None" = None
prefill_chunks: "list[BatchPrefillWithRaggedKVCacheWrapper]" = field(
default_factory=list
)
@@ -495,7 +824,7 @@ def use_flashinfer_prefill() -> bool:
vllm_config = get_current_vllm_config()
if not (
not vllm_config.attention_config.disable_flashinfer_prefill
and flashinfer_available
and flashinfer_available()
and not vllm_config.attention_config.use_cudnn_prefill
and current_platform.is_device_capability_family(100)
):
@@ -509,7 +838,7 @@ def use_cudnn_prefill() -> bool:
vllm_config = get_current_vllm_config()
return (
flashinfer_available
flashinfer_available()
and vllm_config.attention_config.use_cudnn_prefill
and current_platform.is_device_capability_family(100)
and has_nvidia_artifactory()
@@ -731,6 +1060,8 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
has_context = True
if self._fi_prefill_main is None:
from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
self._fi_prefill_main = BatchPrefillWithRaggedKVCacheWrapper(
self._workspace_buffer, "NHD", backend="cutlass"
)
@@ -739,6 +1070,8 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):
num_chunks = chunked_context.cu_seq_lens.shape[0]
# Allocate more prefill chunk wrappers if needed
if len(self._fi_prefill_chunks) < num_chunks:
from flashinfer import BatchPrefillWithRaggedKVCacheWrapper
for _ in range(len(self._fi_prefill_chunks), num_chunks):
self._fi_prefill_chunks.append(
BatchPrefillWithRaggedKVCacheWrapper(
@@ -1513,6 +1846,8 @@ class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]):
):
assert isinstance(prefill, CudnnPrefillMetadata)
assert prefill.query_seq_lens is not None
from flashinfer.prefill import cudnn_batch_prefill_with_kv_cache
output, lse = cudnn_batch_prefill_with_kv_cache(
q=q,
k_cache=k,
@@ -1572,6 +1907,8 @@ class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]):
assert prefill.chunked_context is not None
assert prefill.chunked_context.seq_lens[chunk_idx] is not None
assert prefill.query_seq_lens is not None
from flashinfer.prefill import cudnn_batch_prefill_with_kv_cache
return cudnn_batch_prefill_with_kv_cache(
q=q,
k_cache=k,
@@ -4,11 +4,11 @@ import functools
import torch
from vllm.attention.layer import Attention
from vllm.config import CacheConfig, VllmConfig
from vllm.forward_context import ForwardContext, get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.custom_op import CustomOp
from vllm.model_executor.layers.attention import Attention
from vllm.utils.math_utils import cdiv
from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backend import (
+1 -1
View File
@@ -4,9 +4,9 @@ from dataclasses import dataclass
import torch
from vllm.attention.layer import MLAAttention
from vllm.config import CacheConfig
from vllm.model_executor.custom_op import PluggableLayer
from vllm.model_executor.layers.attention import MLAAttention
from vllm.model_executor.layers.quantization import QuantizationConfig
@@ -19,12 +19,12 @@ from compressed_tensors.quantization import (
from compressed_tensors.transform import TransformConfig
import vllm.envs as envs
from vllm.attention.layer import Attention
from vllm.distributed import (
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import (
LinearBase,
@@ -11,9 +11,9 @@ import vllm.envs as envs
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm import _custom_ops as ops
from vllm._aiter_ops import rocm_aiter_ops
from vllm.attention.layer import Attention
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.batch_invariant import (
vllm_is_batch_invariant,
)
@@ -11,8 +11,8 @@ from torch.nn.parameter import Parameter
import vllm.envs as envs
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
from vllm.attention.layer import Attention
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
@@ -7,9 +7,9 @@ import torch
from torch.nn.parameter import Parameter
from vllm import envs
from vllm.attention.layer import Attention
from vllm.config import get_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import (
FusedMoE,
FusedMoEConfig,
@@ -8,8 +8,8 @@ import regex as re
import torch
from torch.nn.parameter import Parameter
from vllm.attention.layer import Attention
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
LinearBase,
LinearMethodBase,
@@ -7,8 +7,8 @@ import torch
from torch.nn.parameter import Parameter
from vllm import _custom_ops as ops
from vllm.attention.layer import Attention
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
from vllm.model_executor.layers.quantization import QuantizationMethods
from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase
@@ -6,8 +6,8 @@ from typing import TYPE_CHECKING, Any, Optional, cast
import torch
from vllm.attention.layer import Attention
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import (
LinearBase,
+1 -1
View File
@@ -11,9 +11,9 @@ import torch
from torch import nn
from typing_extensions import assert_never
from vllm.attention.layer import Attention, MLAAttention
from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention, MLAAttention
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig,
QuantizeMethodBase,
+1 -1
View File
@@ -9,7 +9,6 @@ from itertools import islice
import torch
from torch import nn
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
@@ -18,6 +17,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -11,7 +11,7 @@ import torch.nn as nn
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.utils import divide
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+2 -2
View File
@@ -32,12 +32,12 @@ import torch
from torch import nn
from transformers import ApertusConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import XIELU
from vllm.model_executor.layers.attention.encoder_only_attention import (
from vllm.model_executor.layers.attention import (
Attention,
EncoderOnlyAttention,
)
from vllm.model_executor.layers.layernorm import RMSNorm
+1 -1
View File
@@ -8,7 +8,6 @@ from itertools import islice
import torch
from torch import nn
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -19,6 +18,7 @@ from vllm.distributed import (
)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -29,7 +29,6 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -38,6 +37,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -32,7 +32,6 @@ import torch.nn.functional as F
from torch import nn
from transformers.configuration_utils import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -41,6 +40,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -9,12 +9,12 @@ import torch
from torch import nn
from transformers import BambaConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -11,7 +11,7 @@ from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, PoolerConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.encoder_only_attention import (
from vllm.model_executor.layers.attention import (
EncoderOnlyAttention,
)
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -15,7 +15,7 @@ from vllm.distributed import (
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.activation import get_act_and_mul_fn, get_act_fn
from vllm.model_executor.layers.attention.encoder_only_attention import (
from vllm.model_executor.layers.attention import (
EncoderOnlyAttention,
)
from vllm.model_executor.layers.fused_moe import activation_without_mul, fused_topk
+1 -1
View File
@@ -11,7 +11,7 @@ from transformers import Blip2VisionConfig, BlipVisionConfig
from vllm.distributed import divide, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
+1 -1
View File
@@ -27,7 +27,6 @@ import torch
from torch import nn
from transformers import BloomConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -36,6 +35,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -16,12 +16,12 @@ from transformers import (
ChameleonVQVAEConfig,
)
from vllm.attention.layer import Attention
from vllm.config import CacheConfig, VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -12,11 +12,11 @@ import torch
from torch import nn
from torch.nn import LayerNorm
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -2
View File
@@ -14,12 +14,11 @@ from transformers import (
CLIPVisionConfig,
)
from vllm.attention.layer import Attention
from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.distributed import divide, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import Attention, MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
+1 -1
View File
@@ -30,11 +30,11 @@ import torch
from torch import nn
from transformers import Cohere2Config, CohereConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -8,13 +8,13 @@ import torch
import torch.nn as nn
from transformers import DbrxConfig
from vllm.attention.layer import Attention
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
get_pp_group,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import (
QKVParallelLinear,
+1 -1
View File
@@ -18,7 +18,7 @@ import torch.nn as nn
import torch.nn.functional as F
from transformers import CLIPVisionConfig
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+1 -1
View File
@@ -33,7 +33,6 @@ from torch import nn
from transformers import DeepseekV2Config, DeepseekV3Config
from vllm._aiter_ops import rocm_aiter_ops
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ParallelConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
@@ -45,6 +44,7 @@ from vllm.distributed import (
)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import LayerNorm, RMSNorm
+1 -1
View File
@@ -32,7 +32,6 @@ import torch
from torch import nn
from transformers import Dots1Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import (
@@ -41,6 +40,7 @@ from vllm.distributed import (
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -16,7 +16,7 @@ from vllm.distributed.parallel_state import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention.mm_encoder_attention import (
from vllm.model_executor.layers.attention import (
MMEncoderAttention,
)
from vllm.model_executor.layers.conv import Conv2dLayer
+1 -1
View File
@@ -32,7 +32,6 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
@@ -42,6 +41,7 @@ from vllm.distributed import (
)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -42,7 +42,7 @@ from vllm.distributed import parallel_state
from vllm.distributed import utils as dist_utils
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import QuickGELU
from vllm.model_executor.layers.attention.mm_encoder_attention import (
from vllm.model_executor.layers.attention import (
MMEncoderAttention,
)
from vllm.model_executor.layers.layernorm import RMSNorm
+1 -2
View File
@@ -31,12 +31,11 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
# from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -32,11 +32,11 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -28,11 +28,11 @@ import torch
from torch import nn
from transformers import Exaone4Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -30,7 +30,6 @@ from torch import nn
from torch.nn import LayerNorm
from transformers import FalconConfig as HF_FalconConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -40,6 +39,7 @@ from vllm.distributed import (
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -9,12 +9,12 @@ import torch
from torch import nn
from transformers import FalconH1Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -26,12 +26,12 @@ import torch
from torch import nn
from transformers import GemmaConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -23,12 +23,12 @@ import torch
from torch import nn
from transformers import Gemma2Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+2 -2
View File
@@ -22,13 +22,13 @@ import torch
from torch import nn
from transformers import Gemma3TextConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.attention.encoder_only_attention import (
from vllm.model_executor.layers.attention import (
Attention,
EncoderOnlyAttention,
)
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
+1 -1
View File
@@ -21,7 +21,6 @@ import torch
from torch import nn
from transformers.models.gemma3n.configuration_gemma3n import Gemma3nTextConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
@@ -32,6 +31,7 @@ from vllm.model_executor.layers.activation import (
GeluAndMul,
GeluAndMulSparse,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
+1 -1
View File
@@ -29,10 +29,10 @@ import torch
from torch import nn
from transformers import Glm4Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear
from vllm.model_executor.layers.logits_processor import LogitsProcessor
+1 -1
View File
@@ -52,7 +52,7 @@ from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
from vllm.distributed import get_tensor_model_parallel_world_size, parallel_state
from vllm.distributed import utils as dist_utils
from vllm.logger import init_logger
from vllm.model_executor.layers.attention.mm_encoder_attention import (
from vllm.model_executor.layers.attention import (
MMEncoderAttention,
)
from vllm.model_executor.layers.conv import Conv2dLayer, Conv3dLayer
+1 -1
View File
@@ -32,7 +32,6 @@ import torch
from torch import nn
from transformers.models.glm4_moe import Glm4MoeConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
@@ -42,6 +41,7 @@ from vllm.distributed import (
)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -23,7 +23,7 @@ from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul, get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
+1 -1
View File
@@ -16,7 +16,7 @@ from vllm.config.multimodal import BaseDummyOptions
from vllm.distributed.parallel_state import get_tensor_model_parallel_world_size
from vllm.inputs.data import PromptType
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -27,7 +27,6 @@ import torch
from torch import nn
from transformers import GPT2Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed.parallel_state import (
@@ -35,6 +34,7 @@ from vllm.distributed.parallel_state import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -28,11 +28,11 @@ import torch
from torch import nn
from transformers import GPTBigCodeConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -26,11 +26,11 @@ import torch
from torch import nn
from transformers import GPTJConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -26,11 +26,11 @@ import torch
from torch import nn
from transformers import GPTNeoXConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -7,7 +7,6 @@ import torch.distributed as dist
from torch import nn
from transformers import GptOssConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -19,6 +18,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.fused_moe.config import FusedMoEParallelConfig
from vllm.model_executor.layers.layernorm import RMSNorm
+1 -1
View File
@@ -31,11 +31,11 @@ import torch
from torch import nn
from transformers import GraniteConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -31,7 +31,6 @@ from typing import Any
import torch
from torch import nn
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -39,6 +38,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
@@ -9,11 +9,11 @@ import torch
from torch import nn
from transformers import GraniteMoeHybridConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear
from vllm.model_executor.layers.logits_processor import LogitsProcessor
+1 -1
View File
@@ -32,12 +32,12 @@ import torch
import torch.nn.functional as F
from torch import nn
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import GeluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -33,7 +33,6 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
@@ -43,6 +42,7 @@ from vllm.distributed import (
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -39,7 +39,7 @@ from vllm.distributed import parallel_state
from vllm.distributed import utils as dist_utils
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
@@ -29,7 +29,7 @@ from transformers.models.idefics2.configuration_idefics2 import (
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
+1 -1
View File
@@ -23,7 +23,7 @@ from vllm.distributed import (
tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -10,7 +10,6 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -21,6 +20,7 @@ from vllm.distributed import (
tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -15,7 +15,7 @@ from transformers import PretrainedConfig
from transformers.utils import torch_int
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear
@@ -24,10 +24,10 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
+1 -1
View File
@@ -20,7 +20,7 @@ from vllm.config import VllmConfig
from vllm.config.model import ModelConfig
from vllm.distributed import parallel_state
from vllm.distributed import utils as dist_utils
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -28,7 +28,6 @@ from itertools import islice
import torch
from torch import nn
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -36,6 +35,7 @@ from vllm.distributed import (
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -31,7 +31,6 @@ import torch
from torch import nn
from transformers import Jais2Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -39,6 +38,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import ReLUSquaredActivation
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+1 -1
View File
@@ -9,11 +9,11 @@ import torch
from torch import nn
from transformers import JambaConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -20,7 +20,7 @@ from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.attention.mm_encoder_attention import (
from vllm.model_executor.layers.attention import (
MMEncoderAttention,
)
from vllm.model_executor.layers.conv import Conv2dLayer
+1 -1
View File
@@ -7,11 +7,11 @@ import torch
import torch.nn as nn
from transformers import Lfm2Config
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear,
+1 -1
View File
@@ -6,7 +6,6 @@ from itertools import islice
import torch
import torch.nn as nn
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config
from vllm.distributed import (
@@ -15,6 +14,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -13,7 +13,7 @@ from transformers import Siglip2VisionConfig
from vllm.compilation.decorators import support_torch_compile
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention
from vllm.model_executor.layers.attention import MMEncoderAttention
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
+2 -2
View File
@@ -31,12 +31,12 @@ import torch
from torch import nn
from transformers import LlamaConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention.encoder_only_attention import (
from vllm.model_executor.layers.attention import (
Attention,
EncoderOnlyAttention,
)
from vllm.model_executor.layers.layernorm import RMSNorm
+2 -2
View File
@@ -24,7 +24,6 @@ import torch
from torch import nn
from transformers import Llama4TextConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -33,7 +32,8 @@ from vllm.distributed import (
tensor_model_parallel_all_gather,
)
from vllm.logger import init_logger
from vllm.model_executor.layers.attention.chunked_local_attention import (
from vllm.model_executor.layers.attention import (
Attention,
ChunkedLocalAttention,
)
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
+1 -1
View File
@@ -6,7 +6,6 @@ from itertools import islice
import torch
from torch import nn
from vllm.attention.layer import Attention
from vllm.config import (
CacheConfig,
VllmConfig,
@@ -22,6 +21,7 @@ from vllm.distributed import (
)
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
+1 -1
View File
@@ -33,7 +33,6 @@ import torch
from torch import nn
from transformers import PretrainedConfig
from vllm.attention.layer import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
@@ -43,6 +42,7 @@ from vllm.distributed import (
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.activation import FatreluAndMul, SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import fused_experts, fused_topk
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (

Some files were not shown because too many files have changed in this diff Show More