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

[Lint]Style: Convert vllm-ascend/ to ruff format(Batch #10) (#6173)

Browse Source 
### What this PR does / why we need it?
**Scope of Changes**:
| File Path |
| :--- |
|`vllm_ascend/ops/layer_shard_linear.py`|
|`vllm_ascend/ops/linear.py`|
|`vllm_ascend/ops/linear_op.py`|
|`vllm_ascend/worker/worker.py`|
| ` vllm_ascend/patch/worker/patch_bert.py` |
| ` vllm_ascend/patch/worker/patch_deepseek.py` |
| ` vllm_ascend/patch/worker/patch_distributed.py` |
| ` vllm_ascend/patch/worker/patch_module.py` |
| ` vllm_ascend/patch/worker/patch_multimodal_merge.py` |
| ` vllm_ascend/patch/worker/patch_qwen3_next.py` |
| ` vllm_ascend/patch/worker/patch_qwen3_next_mtp.py` |
| ` vllm_ascend/patch/worker/patch_rejection_sampler.py` |
| ` vllm_ascend/patch/worker/patch_rope.py` |
| ` vllm_ascend/patch/worker/patch_triton.py` |
| ` vllm_ascend/patch/worker/patch_unquantized_gemm.py` |
| ` vllm_ascend/patch/worker/patch_v2_egale.py` |
|` vllm_ascend/worker/npu_input_batch.py`|
|` vllm_ascend/worker/v2/aclgraph_utils.py`|
|` vllm_ascend/worker/v2/attn_utils.py`|
|` vllm_ascend/worker/v2/model_runner.py`|
|` vllm_ascend/worker/v2/sample/gumbel.py`|
|` vllm_ascend/worker/v2/sample/penalties.py`|
|` vllm_ascend/worker/v2/sample/sampler.py`|
|` vllm_ascend/worker/v2/spec_decode/__init__.py`|
|` vllm_ascend/worker/v2/spec_decode/eagle.py`|
|` vllm_ascend/worker/v2/states.py`|
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.14.0
- vLLM main:
d68209402d

Signed-off-by: MrZ20 <2609716663@qq.com>
Signed-off-by: SILONG ZENG <2609716663@qq.com>
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
SILONG ZENG
2026-02-06 15:35:06 +08:00
committed by GitHub
parent 65b7f716e6
commit 19b5d44ea8
33 changed files with 938 additions and 1243 deletions
Download Patch File Download Diff File Expand all files Collapse all files

104
vllm_ascend/ops/layer_shard_linear.py
View File

@@ -1,6 +1,6 @@
from collections.abc import Callable
from dataclasses import dataclass
from functools import lru_cache
from typing import Callable, List, Optional
import torch
import torch.distributed as dist
@@ -17,39 +17,38 @@ def dispose_tensor(x: torch.Tensor):
@dataclass
class LayerMetadata:
"""Metadata for a layer.
"""
"""Metadata for a layer."""
layer_idx: int # The index of the layer.
layer: LinearBase # The layer object.
post_method: Callable[[
torch.nn.Module
], None] # The `process_weights_after_loading` method from the quant method.
post_method: Callable[[torch.nn.Module], None] # The `process_weights_after_loading` method from the quant method.
weight: torch.Tensor # The weight tensor.
window_idx: int # The index of the window.
@dataclass
class ShardWindowMetadata:
"""Metadata for a shard window.
"""
"""Metadata for a shard window."""
weight: torch.Tensor # The weight tensor to be shard by layers.
data_layer_idx: int # The index of the layer this window's weight is equal to.
work: Optional[torch.distributed.Work] # The asynchronous broadcast work.
work: torch.distributed.Work | None # The asynchronous broadcast work.
@dataclass
class SeriesMetadata:
"""Metadata for a weight shard series.
"""
"""Metadata for a weight shard series."""
group: GroupCoordinator
start_layer: int
end_layer: int
num_layers: int
prefetch_step: int
dummy_weight: torch.Tensor # Dummy weight to replace the loaded weight matrix. All the layers in the series share the same dummy weight tensor.
dummy_weight: torch.Tensor # Dummy weight to replace the loaded weight matrix.
# All the layers in the series share the same dummy weight tensor.
layers: list[LayerMetadata]
shard_windows: list[
ShardWindowMetadata] # Shard windows for prefetching. The window size is (`prefetch_step` + 1), as only the weights for the next (`prefetch_step` + 1) layers need to be stored.
shard_windows: list[ShardWindowMetadata] # Shard windows for prefetching. The window size is (`prefetch_step` + 1),
# as only the weights for the next (`prefetch_step` + 1) layers need to be stored.
window_offset: int # The index of the window for the next coming layer.
def is_source(self, layer_idx) -> bool:
@@ -63,9 +62,9 @@ class SeriesMetadata:
self.layers.sort(key=lambda x: x.layer_idx)
self.num_layers = len(self.layers)
assert self.num_layers > 0, "No layers in the series"
assert self.prefetch_step >= 0 and self.prefetch_step <= max(
0, self.num_layers -
2), "prefetch_step must be in [0, num_layers - 2]"
assert self.prefetch_step >= 0 and self.prefetch_step <= max(0, self.num_layers - 2), (
"prefetch_step must be in [0, num_layers - 2]"
)
self.start_layer = self.layers[0].layer_idx
self.end_layer = self.layers[-1].layer_idx + 1
@@ -73,25 +72,27 @@ class SeriesMetadata:
layer = self.layers[layer_idx - self.start_layer]
assert layer.layer_idx == layer_idx, "layer_idx must be consecutive"
is_source = self.is_source(layer_idx)
# If the weight uses dummy weight, make a copy temporary such that the post method call won't affect other layers which also uses dummy weight.
# If the weight uses dummy weight, make a copy temporary such that the post method call
# won't affect other layers which also uses dummy weight.
if not is_source:
layer.weight.set_(torch.empty_like(self.dummy_weight))
# Broadcast to get the true weight.
dist.broadcast(layer.weight,
src=self.group.ranks[layer_idx %
self.group.world_size],
group=self.group.device_group)
dist.broadcast(
layer.weight, src=self.group.ranks[layer_idx % self.group.world_size], group=self.group.device_group
)
# Call `process_weights_after_loading` from the quant method.
layer.post_method(layer.layer)
step = layer_idx - self.start_layer
if step < self.prefetch_step:
# Build the windows for the first `prefetch_step` layers. The weights can be used for the first `prefetch_step` layers in `forward()`, so also clone the weights.
# Build the windows for the first `prefetch_step` layers. The weights can be used
# for the first `prefetch_step` layers in `forward()`, so also clone the weights.
self.shard_windows.append(
ShardWindowMetadata(
weight=layer.weight.clone().detach(),
data_layer_idx=layer_idx,
work=None,
))
)
)
layer.window_idx = step
# When the layer not intended to be stored in this device, link to the corresponding window's tensor.
if not is_source:
@@ -104,7 +105,8 @@ class SeriesMetadata:
weight=torch.empty_like(layer.weight),
data_layer_idx=-1,
work=None,
))
)
)
# When the layer not intended to be stored in this device, dispose the tensor.
if not is_source:
dispose_tensor(layer.weight)
@@ -113,8 +115,7 @@ class SeriesMetadata:
def reach_layer(self, layer_idx: int):
# The index of the layer to be prefetched.
next_layer_idx = (layer_idx + self.prefetch_step
) % self.num_layers + self.start_layer
next_layer_idx = (layer_idx + self.prefetch_step) % self.num_layers + self.start_layer
next_layer = self.layers[next_layer_idx - self.start_layer]
# The index of the window to store the weight for the coming layer.
next_layer.window_idx = self.window_offset
@@ -123,8 +124,7 @@ class SeriesMetadata:
if not self.is_source(next_layer_idx):
next_layer.weight.set_(window.weight)
# Update `window_offset` by rolling one step.
self.window_offset = (self.window_offset + 1) % (self.prefetch_step +
1)
self.window_offset = (self.window_offset + 1) % (self.prefetch_step + 1)
assert window.data_layer_idx != next_layer_idx
window.data_layer_idx = next_layer_idx
# Start asynchronous broadcast work.
@@ -132,13 +132,13 @@ class SeriesMetadata:
next_layer.weight,
src=self.group.ranks[next_layer_idx % self.group.world_size],
group=self.group.device_group,
async_op=True)
async_op=True,
)
def wait_weight(self, layer_idx: int):
# Find the asynchronous broadcast work and wait for it.
assert self.shard_windows
window = self.shard_windows[self.layers[layer_idx -
self.start_layer].window_idx]
window = self.shard_windows[self.layers[layer_idx - self.start_layer].window_idx]
# Make sure the data in the corresponding shard window is for the current layer.
assert window.data_layer_idx == layer_idx
if window.work is not None:
@@ -148,8 +148,8 @@ class SeriesMetadata:
@dataclass
class LayerExternalMetadata:
"""External metadata for a layer.
"""
"""External metadata for a layer."""
series: SeriesMetadata
layer_idx: int
@@ -159,9 +159,7 @@ _series_dict: dict[str, SeriesMetadata] = {}
_layer_external_dict: dict[int, LayerExternalMetadata] = {}
def _create_forward_wrapper(forward: Callable, series: SeriesMetadata,
layer_idx: int) -> Callable:
def _create_forward_wrapper(forward: Callable, series: SeriesMetadata, layer_idx: int) -> Callable:
def wrapped_forward(*args, **kwargs):
# Wait for the weight.
series.wait_weight(layer_idx)
@@ -173,23 +171,32 @@ def _create_forward_wrapper(forward: Callable, series: SeriesMetadata,
"""
Register linear layers into a shard storage series.
In a parallel group, each device stores a distinct, non-overlapping subset of layers from the series. All layers in a series must have the same structure (are isomorphic). The weight matrix for the i-th layer is stored on device (i % n), where n is the number of devices.
In a parallel group, each device stores a distinct, non-overlapping subset of layers from the series.
All layers in a series must have the same structure (are isomorphic). The weight matrix for the i-th layer
is stored on device (i % n), where n is the number of devices.
After loading the model, you must call `post_process_after_loading_for_shard_weight_series(layer)` on any layer of this series to complete the initialization.
After loading the model, you must call `post_process_after_loading_for_shard_weight_series(layer)`
on any layer of this series to complete the initialization.
During execution, each time a new layer is reached, you must call `reach_layer_for_shard_weight_series(layer)` for that layer to prefetch the weights. The argument `prefetch_step` is a non-negative integer k that manages asynchronous weight prefetching. Each call to `reach_layer_for_shard_weight_series(current_layer)` method will trigger an asynchronous prefetch for the weights of the k-th subsequent layer after `current_layer` within the series.
During execution, each time a new layer is reached, you must call `reach_layer_for_shard_weight_series(layer)`
for that layer to prefetch the weights. The argument `prefetch_step` is a non-negative integer k that manages
asynchronous weight prefetching. Each call to `reach_layer_for_shard_weight_series(current_layer)` method will
trigger an asynchronous prefetch for the weights of the k-th subsequent layer after `current_layer` within the series.
Note: The layers are managed as a circular buffer. The index of the layer to prefetch is determined by the formula:
- start_layer is the index of the first layer in the series (inclusive).
- end_layer is the index of the last layer in the series (exclusive). Thus, the series includes all layers with indices in the range [start_layer, end_layer).
- end_layer is the index of the last layer in the series (exclusive). Thus, the series includes all layers with
indices in the range [start_layer, end_layer).
- total_layers = end_layer - start_layer
- prefetch_layer_idx = (layer_idx + prefetch_step) % total_layers + start_layer
To hold the weights for the current layer and the k prefetched layers, a pool of (k + 1) shard tensor buffers will be created for this series.
To hold the weights for the current layer and the k prefetched layers, a pool of (k + 1) shard tensor buffers
will be created for this series.
Arguments:
series_name: This name identifies which series this layer belongs to.
group: The group coordinator for handling asynchronous communications. It is recommended to create a new group coordinator for each new series.
group: The group coordinator for handling asynchronous communications. It is recommended to create a new group
coordinator for each new series.
layer: The linear layer object to register.
prefetch_step: An integer that manages asynchronous weight prefetching. Setting it to 0 or 1 can cover most cases.
"""
@@ -224,7 +231,8 @@ def register_layer_to_shard_weight_series(
post_method=layer.quant_method.process_weights_after_loading,
weight=layer.weight,
window_idx=-1,
))
)
)
# Discard the original `process_weights_after_loading` method such that it won't be called by others.
layer.quant_method.process_weights_after_loading = lambda layer: None
# When the layer not intended to be stored in this device, dispose the tensor and skip weight loading.
@@ -257,6 +265,7 @@ def wait_layer_for_shard_weight_series(layer: LinearBase):
@lru_cache(maxsize=1)
def get_current_model_num_hidden_layers() -> int:
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
return vllm_config.model_config.get_total_num_hidden_layers()
@@ -268,10 +277,11 @@ def is_hidden_layer(layer: LinearBase) -> bool:
def register_all_layers_to_shard_weight_series(
layer_sharding: List[LinearBase], ):
for curr_layer in (layer_sharding or []):
layer_sharding: list[LinearBase],
):
for curr_layer in layer_sharding or []:
if is_hidden_layer(curr_layer):
layer_name = curr_layer.prefix.split('.')[-1]
layer_name = curr_layer.prefix.split(".")[-1]
register_layer_to_shard_weight_series(
series_name=layer_name,
group=get_shard_weight_group(),

285
vllm_ascend/ops/linear.py
View File

@@ -20,19 +20,23 @@ AscendMergedColumnParallelLinear, AscendMergedColumnParallelLinear,
AscendRowParallelLinear and AscendColumnParallelLinear.
"""
from typing import Optional, Union
import torch
import torch.nn as nn
from torch.nn.parameter import Parameter
from vllm.config import get_current_vllm_config
from vllm.distributed import divide
from vllm.model_executor.layers.linear import ( # noqa
WEIGHT_LOADER_V2_SUPPORTED, ColumnParallelLinear, LinearBase,
MergedColumnParallelLinear, QKVParallelLinear, QuantizeMethodBase,
ReplicatedLinear, RowParallelLinear, UnquantizedLinearMethod)
from vllm.model_executor.layers.quantization.base_config import \
QuantizationConfig
WEIGHT_LOADER_V2_SUPPORTED,
ColumnParallelLinear,
LinearBase,
MergedColumnParallelLinear,
QKVParallelLinear,
QuantizeMethodBase,
ReplicatedLinear,
RowParallelLinear,
UnquantizedLinearMethod,
)
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
from vllm.model_executor.utils import set_weight_attrs
from vllm_ascend.ops.linear_op import get_parallel_op, get_replicated_op
@@ -50,14 +54,13 @@ class AscendUnquantizedLinearMethod(UnquantizedLinearMethod):
# TODO(realliujiaxu): Remove this class after linear of vllm supports custom comm group
class AscendLinearBase(LinearBase):
def __init__(
self,
input_size: int,
output_size: int,
skip_bias_add: bool = False,
params_dtype: Optional[torch.dtype] = None,
quant_config: Optional[QuantizationConfig] = None,
params_dtype: torch.dtype | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
*,
return_bias: bool = True,
@@ -75,11 +78,9 @@ class AscendLinearBase(LinearBase):
self.quant_config = quant_config
self.prefix = prefix
if quant_config is None:
self.quant_method: Optional[
QuantizeMethodBase] = AscendUnquantizedLinearMethod()
self.quant_method: QuantizeMethodBase | None = AscendUnquantizedLinearMethod()
else:
self.quant_method = quant_config.get_quant_method(self,
prefix=prefix)
self.quant_method = quant_config.get_quant_method(self, prefix=prefix)
self.return_bias = return_bias
self.disable_tp = disable_tp
@@ -100,11 +101,11 @@ class AscendQKVParallelLinear(QKVParallelLinear):
hidden_size: int,
head_size: int,
total_num_heads: int,
total_num_kv_heads: Optional[int] = None,
total_num_kv_heads: int | None = None,
bias: bool = True,
skip_bias_add: bool = False,
params_dtype: Optional[torch.dtype] = None,
quant_config: Optional[QuantizationConfig] = None,
params_dtype: torch.dtype | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
*,
return_bias: bool = True,
@@ -112,9 +113,9 @@ class AscendQKVParallelLinear(QKVParallelLinear):
v_head_size: int | None = None,
):
self.v_head_size = v_head_size if v_head_size is not None else head_size
self.custom_op, _, tp_size = get_parallel_op(disable_tp, prefix, self,
"column")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after linear of vllm supports custom comm group
self.custom_op, _, tp_size = get_parallel_op(disable_tp, prefix, self, "column")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after
# linear of vllm supports custom comm group
self.hidden_size = hidden_size
self.head_size = head_size
self.total_num_heads = total_num_heads
@@ -125,35 +126,35 @@ class AscendQKVParallelLinear(QKVParallelLinear):
self.num_heads = divide(self.total_num_heads, tp_size)
if tp_size >= self.total_num_kv_heads:
self.num_kv_heads = 1
self.num_kv_head_replicas = divide(tp_size,
self.total_num_kv_heads)
self.num_kv_head_replicas = divide(tp_size, self.total_num_kv_heads)
else:
self.num_kv_heads = divide(self.total_num_kv_heads, tp_size)
self.num_kv_head_replicas = 1
input_size = self.hidden_size
output_size = (self.num_heads +
2 * self.num_kv_heads) * tp_size * self.head_size
output_size = (self.num_heads + 2 * self.num_kv_heads) * tp_size * self.head_size
self.output_sizes = [
self.num_heads * self.head_size * tp_size, # q_proj
self.num_kv_heads * self.head_size * tp_size, # k_proj
self.num_kv_heads * self.head_size * tp_size, # v_proj
]
AscendColumnParallelLinear.__init__(self,
input_size=input_size,
output_size=output_size,
bias=bias,
gather_output=False,
skip_bias_add=skip_bias_add,
params_dtype=params_dtype,
quant_config=quant_config,
prefix=prefix,
return_bias=return_bias,
disable_tp=disable_tp)
AscendColumnParallelLinear.__init__(
self,
input_size=input_size,
output_size=output_size,
bias=bias,
gather_output=False,
skip_bias_add=skip_bias_add,
params_dtype=params_dtype,
quant_config=quant_config,
prefix=prefix,
return_bias=return_bias,
disable_tp=disable_tp,
)
def forward(
self,
input_,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.custom_op is not None:
return self.custom_op.apply(input_)
@@ -178,35 +179,36 @@ class AscendMergedColumnParallelLinear(MergedColumnParallelLinear):
bias: bool = True,
gather_output: bool = False,
skip_bias_add: bool = False,
params_dtype: Optional[torch.dtype] = None,
quant_config: Optional[QuantizationConfig] = None,
params_dtype: torch.dtype | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
*,
return_bias: bool = True,
disable_tp: bool = False,
):
self.custom_op, self.tp_rank, self.tp_size = get_parallel_op(
disable_tp, prefix, self, "column")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after linear of vllm supports custom comm group
self.custom_op, self.tp_rank, self.tp_size = get_parallel_op(disable_tp, prefix, self, "column")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after
# linear of vllm supports custom comm group
self.output_sizes = output_sizes
assert all(output_size % self.tp_size == 0
for output_size in output_sizes)
AscendColumnParallelLinear.__init__(self,
input_size=input_size,
output_size=sum(output_sizes),
bias=bias,
gather_output=gather_output,
skip_bias_add=skip_bias_add,
params_dtype=params_dtype,
quant_config=quant_config,
prefix=prefix,
return_bias=return_bias,
disable_tp=disable_tp)
assert all(output_size % self.tp_size == 0 for output_size in output_sizes)
AscendColumnParallelLinear.__init__(
self,
input_size=input_size,
output_size=sum(output_sizes),
bias=bias,
gather_output=gather_output,
skip_bias_add=skip_bias_add,
params_dtype=params_dtype,
quant_config=quant_config,
prefix=prefix,
return_bias=return_bias,
disable_tp=disable_tp,
)
def forward(
self,
input_,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.custom_op is not None:
return self.custom_op.apply(input_)
@@ -229,9 +231,9 @@ class AscendRowParallelLinear(RowParallelLinear):
bias: bool = True,
input_is_parallel: bool = True,
skip_bias_add: bool = False,
params_dtype: Optional[torch.dtype] = None,
params_dtype: torch.dtype | None = None,
reduce_results: bool = True,
quant_config: Optional[QuantizationConfig] = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
*,
return_bias: bool = True,
@@ -247,23 +249,25 @@ class AscendRowParallelLinear(RowParallelLinear):
self.unique_prefix = unique_prefix
compilation_config.static_forward_context[unique_prefix] = self
self.custom_op, self.tp_rank, self.tp_size = get_parallel_op(
disable_tp, prefix, self, "row")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after linear of vllm supports custom comm group
self.custom_op, self.tp_rank, self.tp_size = get_parallel_op(disable_tp, prefix, self, "row")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after
# linear of vllm supports custom comm group
# Divide the weight matrix along the first dimension.
self.input_size_per_partition = divide(input_size, self.tp_size)
self.output_size_per_partition = output_size
self.output_partition_sizes = [output_size]
AscendLinearBase.__init__(self,
input_size,
output_size,
skip_bias_add,
params_dtype,
quant_config,
prefix,
return_bias=return_bias,
disable_tp=disable_tp)
AscendLinearBase.__init__(
self,
input_size,
output_size,
skip_bias_add,
params_dtype,
quant_config,
prefix,
return_bias=return_bias,
disable_tp=disable_tp,
)
self.input_is_parallel = input_is_parallel
self.reduce_results = reduce_results
@@ -277,19 +281,23 @@ class AscendRowParallelLinear(RowParallelLinear):
output_size=self.output_size,
params_dtype=self.params_dtype,
weight_loader=(
self.weight_loader_v2 if self.quant_method.__class__.__name__
in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
self.weight_loader_v2
if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED
else self.weight_loader
),
)
if not reduce_results and (bias and not skip_bias_add):
raise ValueError("When not reduce the results, adding bias to the "
"results can lead to incorrect results")
raise ValueError("When not reduce the results, adding bias to the results can lead to incorrect results")
if bias:
self.bias = Parameter(
torch.empty(self.output_size, dtype=params_dtype))
set_weight_attrs(self.bias, {
"output_dim": 0,
"weight_loader": self.weight_loader,
})
self.bias = Parameter(torch.empty(self.output_size, dtype=params_dtype))
set_weight_attrs(
self.bias,
{
"output_dim": 0,
"weight_loader": self.weight_loader,
},
)
else:
self.register_parameter("bias", None)
@@ -300,7 +308,7 @@ class AscendRowParallelLinear(RowParallelLinear):
self,
input_,
**kwargs,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.custom_op is not None:
return self.custom_op.apply(input_)
@@ -321,36 +329,36 @@ class AscendColumnParallelLinear(ColumnParallelLinear):
bias: bool = True,
gather_output: bool = False,
skip_bias_add: bool = False,
params_dtype: Optional[torch.dtype] = None,
quant_config: Optional[QuantizationConfig] = None,
output_sizes: Optional[list[int]] = None,
params_dtype: torch.dtype | None = None,
quant_config: QuantizationConfig | None = None,
output_sizes: list[int] | None = None,
prefix: str = "",
*,
return_bias: bool = True,
disable_tp: bool = False,
):
self.custom_op, self.tp_rank, self.tp_size = get_parallel_op(
disable_tp, prefix, self, "column")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after linear of vllm supports custom comm group
#
self.custom_op, self.tp_rank, self.tp_size = get_parallel_op(disable_tp, prefix, self, "column")
# TODO(realliujiaxu): Replace the initialization code below with super().__init__ after
# linear of vllm supports custom comm group
self.input_size_per_partition = input_size
self.output_size_per_partition = divide(output_size, self.tp_size)
self.output_partition_sizes = [self.output_size_per_partition]
# If QKV or MergedColumn, use output size of each partition.
if hasattr(self, "output_sizes"):
self.output_partition_sizes = [
divide(output_size, self.tp_size)
for output_size in self.output_sizes
]
self.output_partition_sizes = [divide(output_size, self.tp_size) for output_size in self.output_sizes]
AscendLinearBase.__init__(self,
input_size,
output_size,
skip_bias_add,
params_dtype,
quant_config,
prefix,
return_bias=return_bias,
disable_tp=disable_tp)
AscendLinearBase.__init__(
self,
input_size,
output_size,
skip_bias_add,
params_dtype,
quant_config,
prefix,
return_bias=return_bias,
disable_tp=disable_tp,
)
self.gather_output = gather_output
@@ -366,16 +374,20 @@ class AscendColumnParallelLinear(ColumnParallelLinear):
output_size=self.output_size,
params_dtype=self.params_dtype,
weight_loader=(
self.weight_loader_v2 if self.quant_method.__class__.__name__
in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
self.weight_loader_v2
if self.quant_method.__class__.__name__ in WEIGHT_LOADER_V2_SUPPORTED
else self.weight_loader
),
)
if bias:
self.bias = Parameter(
torch.empty(self.output_size_per_partition,
dtype=params_dtype))
set_weight_attrs(self.bias, {
"output_dim": 0,
"weight_loader": self.weight_loader,
})
self.bias = Parameter(torch.empty(self.output_size_per_partition, dtype=params_dtype))
set_weight_attrs(
self.bias,
{
"output_dim": 0,
"weight_loader": self.weight_loader,
},
)
else:
self.register_parameter("bias", None)
@@ -385,7 +397,7 @@ class AscendColumnParallelLinear(ColumnParallelLinear):
def forward(
self,
input_,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.custom_op is not None:
return self.custom_op.apply(input_)
@@ -414,8 +426,8 @@ class AscendReplicatedLinear(ReplicatedLinear):
output_size: int,
bias: bool = True,
skip_bias_add: bool = False,
params_dtype: Optional[torch.dtype] = None,
quant_config: Optional[QuantizationConfig] = None,
params_dtype: torch.dtype | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
*,
return_bias: bool = True,
@@ -428,32 +440,39 @@ class AscendReplicatedLinear(ReplicatedLinear):
else:
self.output_partition_sizes = [output_size]
AscendLinearBase.__init__(self,
input_size,
output_size,
skip_bias_add,
params_dtype,
quant_config,
prefix=prefix,
return_bias=return_bias,
disable_tp=disable_tp)
AscendLinearBase.__init__(
self,
input_size,
output_size,
skip_bias_add,
params_dtype,
quant_config,
prefix=prefix,
return_bias=return_bias,
disable_tp=disable_tp,
)
# All the linear layer supports quant method.
assert self.quant_method is not None
self.quant_method.create_weights(self,
self.input_size, [self.output_size],
self.input_size,
self.output_size,
self.params_dtype,
weight_loader=self.weight_loader)
self.quant_method.create_weights(
self,
self.input_size,
[self.output_size],
self.input_size,
self.output_size,
self.params_dtype,
weight_loader=self.weight_loader,
)
if bias:
self.bias = Parameter(
torch.empty(self.output_size, dtype=self.params_dtype))
set_weight_attrs(self.bias, {
"output_dim": 0,
"weight_loader": self.weight_loader,
})
self.bias = Parameter(torch.empty(self.output_size, dtype=self.params_dtype))
set_weight_attrs(
self.bias,
{
"output_dim": 0,
"weight_loader": self.weight_loader,
},
)
else:
self.register_parameter("bias", None)
@@ -463,7 +482,7 @@ class AscendReplicatedLinear(ReplicatedLinear):
def forward(
self,
input_,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.custom_op is not None:
return self.custom_op.apply(input_)

310
vllm_ascend/ops/linear_op.py
View File

@@ -31,16 +31,18 @@ CustomLinearOp
└── CustomReplicatedOp
How to extend a new linear op? Taking column parallel op as an example:
1. Inherit from CustomColumnParallelOp and create a new class MyColumnParallelOp
2. [Optional] The default communication group is the TP group. If a custom communication group is needed, override the comm_group method
2. [Optional] The default communication group is the TP group. If a custom communication group is needed,
override the comm_group method
3. Override the apply method according to requirements, which will replace the original linear.forward
4. Add selection logic for MyColumnParallelOp in the get_column_parallel_op method, typically based on prefix and configuration judgments
Row parallel op follows a similar approach - inherit from RowColumnParallelOp and register the new class in get_row_parallel_op.
4. Add selection logic for MyColumnParallelOp in the get_column_parallel_op method, typically based on
prefix and configuration judgments
Row parallel op follows a similar approach - inherit from RowColumnParallelOp and register the new class in
get_row_parallel_op.
"""
import re
from functools import lru_cache
from types import SimpleNamespace
from typing import Optional, Union
import torch
import torch.distributed as dist
@@ -49,27 +51,37 @@ import torch_npu
from torch import nn
from torch.distributed import ProcessGroup
from torch.nn.parameter import Parameter
from vllm.distributed import (split_tensor_along_last_dim,
tensor_model_parallel_all_reduce,
tensor_model_parallel_reduce_scatter)
from vllm.distributed import (
split_tensor_along_last_dim,
tensor_model_parallel_all_reduce,
tensor_model_parallel_reduce_scatter,
)
from vllm.distributed.parallel_state import get_tp_group
from vllm.forward_context import get_forward_context
from vllm_ascend import envs as envs_ascend
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.distributed.parallel_state import (get_flashcomm2_odp_group,
get_flashcomm2_otp_group,
get_mlp_tp_group,
get_otp_group)
from vllm_ascend.distributed.parallel_state import (
get_flashcomm2_odp_group,
get_flashcomm2_otp_group,
get_mlp_tp_group,
get_otp_group,
)
from vllm_ascend.ops.flashcomm2_oshard_manager import flashcomm2_oshard_manager
from vllm_ascend.utils import (enable_dsa_cp, enable_dsa_cp_with_layer_shard, enable_sp, flashcomm2_enable,
get_flashcomm2_reorgnized_batch_ids,
matmul_allreduce_enable, mlp_tp_enable,
oproj_tp_enable, shared_expert_dp_enabled,
get_weight_prefetch_method)
from vllm_ascend.utils import (
enable_dsa_cp,
enable_dsa_cp_with_layer_shard,
enable_sp,
flashcomm2_enable,
get_flashcomm2_reorgnized_batch_ids,
get_weight_prefetch_method,
matmul_allreduce_enable,
mlp_tp_enable,
oproj_tp_enable,
shared_expert_dp_enabled,
)
class CustomLinearOp:
def __init__(self, layer):
self.layer = layer
self.bias = None
@@ -112,7 +124,6 @@ class CustomLinearOp:
class CustomColumnParallelOp(CustomLinearOp):
def __init__(self, layer):
super().__init__(layer)
self.gather_output = None
@@ -123,7 +134,6 @@ class CustomColumnParallelOp(CustomLinearOp):
class CustomRowParallelOp(CustomLinearOp):
def __init__(self, layer):
super().__init__(layer)
self.reduce_results = None
@@ -140,7 +150,9 @@ class CustomRowParallelOp(CustomLinearOp):
output, output_bias = self.apply_impl(input_)
weight_prefetch_method = get_weight_prefetch_method()
if weight_prefetch_method:
weight_prefetch_method.maybe_prefetch_mlp_weight_preprocess(weight_prefetch_method.MLP_GATE_UP, output, self.prefix)
weight_prefetch_method.maybe_prefetch_mlp_weight_preprocess(
weight_prefetch_method.MLP_GATE_UP, output, self.prefix
)
if not self.return_bias:
return output
@@ -148,7 +160,6 @@ class CustomRowParallelOp(CustomLinearOp):
class CustomReplicatedOp(CustomLinearOp):
def apply_impl(self, input_):
bias = self.bias if not self.skip_bias_add else None
assert self.quant_method is not None
@@ -160,7 +171,6 @@ class CustomReplicatedOp(CustomLinearOp):
class MLPColumnParallelOp(CustomColumnParallelOp):
def __init__(self, layer):
super().__init__(layer)
@@ -171,7 +181,7 @@ class MLPColumnParallelOp(CustomColumnParallelOp):
def apply_impl(
self,
input_: torch.Tensor,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
bias = self.bias if not self.skip_bias_add else None
# Matrix multiply.
assert self.quant_method is not None
@@ -183,7 +193,6 @@ class MLPColumnParallelOp(CustomColumnParallelOp):
class MLPRowParallelOp(CustomRowParallelOp):
def __init__(self, layer):
super().__init__(layer)
@@ -191,22 +200,16 @@ class MLPRowParallelOp(CustomRowParallelOp):
def comm_group(self):
return get_mlp_tp_group()
def apply_impl(
self, input_: torch.Tensor
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
def apply_impl(self, input_: torch.Tensor) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.input_is_parallel:
input_parallel = input_
else:
splitted_input = split_tensor_along_last_dim(
input_, num_partitions=self.tp_size)
splitted_input = split_tensor_along_last_dim(input_, num_partitions=self.tp_size)
input_parallel = splitted_input[self.tp_rank].contiguous()
assert self.quant_method is not None
bias_ = None if (self.tp_rank > 0
or self.skip_bias_add) else self.layer.bias
output_parallel = self.quant_method.apply(self.layer,
input_parallel,
bias=bias_)
bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.layer.bias
output_parallel = self.quant_method.apply(self.layer, input_parallel, bias=bias_)
output = self.comm_group.reduce_scatter(output_parallel, 0)
output_bias = self.bias if self.skip_bias_add else None
@@ -214,7 +217,6 @@ class MLPRowParallelOp(CustomRowParallelOp):
class OProjRowParallelOp(CustomRowParallelOp):
def __init__(self, layer):
super().__init__(layer)
@@ -225,13 +227,11 @@ class OProjRowParallelOp(CustomRowParallelOp):
def apply_impl(
self,
input_: torch.Tensor,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.input_is_parallel:
input_parallel = input_
else:
splitted_input = split_tensor_along_last_dim(
input_, num_partitions=self.tp_size)
splitted_input = split_tensor_along_last_dim(input_, num_partitions=self.tp_size)
input_parallel = splitted_input[self.tp_rank].contiguous()
# Prepare tensors for all-to-all communication
@@ -241,27 +241,19 @@ class OProjRowParallelOp(CustomRowParallelOp):
# Reshape tensor for efficient cross-device transfer:
# [batch, dim] -> [tp_size, batch, chunk] -> flattened
send_buf = (input_parallel.reshape(-1,
self.tp_size, chunk_size).transpose(
0, 1).contiguous().view(-1))
send_buf = input_parallel.reshape(-1, self.tp_size, chunk_size).transpose(0, 1).contiguous().view(-1)
# Create receive buffer
recv_buf = torch.empty(total_batch_size * chunk_size,
dtype=input_parallel.dtype,
device=input_parallel.device)
recv_buf = torch.empty(total_batch_size * chunk_size, dtype=input_parallel.dtype, device=input_parallel.device)
# Perform all-to-all communication
dist.all_to_all_single(recv_buf,
send_buf,
group=self.comm_group.device_group)
dist.all_to_all_single(recv_buf, send_buf, group=self.comm_group.device_group)
input_parallel = recv_buf.view(total_batch_size, chunk_size)
# Only fuse bias add for rank 0 to avoid duplicate bias addition in TP>1
bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
assert self.quant_method is not None
output_parallel = self.quant_method.apply(self.layer,
input_parallel,
bias=bias_)
output_parallel = self.quant_method.apply(self.layer, input_parallel, bias=bias_)
# otp-specific: Combine partial results across devices
output = self.comm_group.reduce_scatter(output_parallel, dim=0)
@@ -278,14 +270,12 @@ class OProjRowParallelOp(CustomRowParallelOp):
class Flashcomm2OProjRowParallelOp(CustomRowParallelOp):
def __init__(self, layer):
super().__init__(layer)
self.odp_group = get_flashcomm2_odp_group()
self.odp_size = self.odp_group.world_size
self.otp_size = get_ascend_config().flashcomm2_oproj_tensor_parallel_size
self.reorgnized_batch_ids = get_flashcomm2_reorgnized_batch_ids(
get_tp_group().world_size)
self.reorgnized_batch_ids = get_flashcomm2_reorgnized_batch_ids(get_tp_group().world_size)
self.group_indices = torch.tensor(self.reorgnized_batch_ids).npu()
self.layer._quant_comm_config = {}
@@ -308,32 +298,28 @@ class Flashcomm2OProjRowParallelOp(CustomRowParallelOp):
def apply_impl(
self,
input_: torch.Tensor,
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
"""Linear layer for Flashcomm2.
Input.ahspe = [batchsize*seqlength, headnum*headdim/TP]
Output.shape = [(batchsize*seqlength+padsize)/TP, hiddensize]
Input.ahspe = [batchsize*seqlength, headnum*headdim/TP]
Output.shape = [(batchsize*seqlength+padsize)/TP, hiddensize]
"""
# Handle input parallelism - split or use as-is
if self.input_is_parallel:
input_parallel = input_
else:
tp_rank = self.tp_rank
splitted_input = split_tensor_along_last_dim(
input_, num_partitions=self.tp_size)
splitted_input = split_tensor_along_last_dim(input_, num_partitions=self.tp_size)
input_parallel = splitted_input[tp_rank].contiguous()
# padding for all-to-all
forward_context = get_forward_context()
num_padding_tokens = forward_context.pad_size
if num_padding_tokens > 0:
input_parallel = nn.functional.pad(input_parallel,
(0, 0, 0, num_padding_tokens))
input_parallel = nn.functional.pad(input_parallel, (0, 0, 0, num_padding_tokens))
def otp_maybe_quant_comm(x):
# Reorganize the tensor so that the batch id and rank id correspond to each other.
chunk_num = len(self.reorgnized_batch_ids) * len(
self.reorgnized_batch_ids[0])
chunk_num = len(self.reorgnized_batch_ids) * len(self.reorgnized_batch_ids[0])
batch_size = x.size(0)
assert batch_size % chunk_num == 0, f"Batch_size({batch_size}) must be divisible by chunk_num({chunk_num})"
@@ -352,26 +338,19 @@ class Flashcomm2OProjRowParallelOp(CustomRowParallelOp):
total_intermediate_size = local_intermediate_size * all2all_tp_size
# Create receive buffer
recv_buf = torch.empty(total_intermediate_size * chunk_size,
dtype=x.dtype,
device=x.device)
recv_buf = torch.empty(total_intermediate_size * chunk_size, dtype=x.dtype, device=x.device)
# Perform all-to-all communication
dist.all_to_all_single(recv_buf,
send_buf,
group=self.odp_group.device_group)
dist.all_to_all_single(recv_buf, send_buf, group=self.odp_group.device_group)
return recv_buf.view(all2all_tp_size, chunk_size,
-1).transpose(0, 1).reshape(chunk_size, -1)
return recv_buf.view(all2all_tp_size, chunk_size, -1).transpose(0, 1).reshape(chunk_size, -1)
if not hasattr(self, "_quant_comm_config"):
self.layer._quant_comm_config = {}
self.layer._quant_comm_config[
"communication_fn"] = otp_maybe_quant_comm
actual_quant_method = getattr(self.quant_method, 'quant_method',
self.quant_method)
from vllm_ascend.quantization.methods.w8a8_static import \
AscendW8A8LinearMethod
self.layer._quant_comm_config["communication_fn"] = otp_maybe_quant_comm
actual_quant_method = getattr(self.quant_method, "quant_method", self.quant_method)
from vllm_ascend.quantization.methods.w8a8_static import AscendW8A8LinearMethod
if not isinstance(actual_quant_method, AscendW8A8LinearMethod):
# Check if w8a8 quantization is enabled. If not, communicate immediately.
input_parallel = otp_maybe_quant_comm(input_parallel)
@@ -382,9 +361,7 @@ class Flashcomm2OProjRowParallelOp(CustomRowParallelOp):
# bias will not get added more than once in TP>1 case)
bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
output_parallel = self.quant_method.apply(self.layer,
input_parallel,
bias=bias_)
output_parallel = self.quant_method.apply(self.layer, input_parallel, bias=bias_)
# output_parallel shape: [bs/(TP/flashcomm2_otp_size), hiddenstate]
if self.tp_size > 1:
# flashcomm2 with reduce-scatter
@@ -408,8 +385,7 @@ class Flashcomm2OProjRowParallelOp(CustomRowParallelOp):
self.input_is_parallel = self.layer.input_is_parallel
self.input_size_per_partition = self.layer.input_size_per_partition
if flashcomm2_oshard_manager.flashcomm2_oshard_enable():
flashcomm2_oshard_manager.register_layer(self.layer,
prefetch_step=1)
flashcomm2_oshard_manager.register_layer(self.layer, prefetch_step=1)
class MatmulAllreduceRowParallelOp(CustomRowParallelOp):
@@ -419,28 +395,22 @@ class MatmulAllreduceRowParallelOp(CustomRowParallelOp):
super().__init__(layer)
self.hcomm_info = self.get_hcomm_info(self.comm_group.device_group)
def apply_impl(
self, input_: torch.Tensor
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
def apply_impl(self, input_: torch.Tensor) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
if self.input_is_parallel:
input_parallel = input_
else:
splitted_input = split_tensor_along_last_dim(
input_, num_partitions=self.tp_size)
splitted_input = split_tensor_along_last_dim(input_, num_partitions=self.tp_size)
input_parallel = splitted_input[self.tp_rank].contiguous()
"""Calculate the output tensor of forward by considering
fusing communication and computation."""
bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
if self.reduce_results and self.tp_size > 1:
output = torch_npu.npu_mm_all_reduce_base(input_parallel,
self.layer.weight.t(),
self.hcomm_info,
bias=bias_)
output = torch_npu.npu_mm_all_reduce_base(
input_parallel, self.layer.weight.t(), self.hcomm_info, bias=bias_
)
else:
assert self.quant_method is not None
output = self.quant_method.apply(self.layer,
input_parallel,
bias=bias_)
output = self.quant_method.apply(self.layer, input_parallel, bias=bias_)
output_bias = self.bias if self.skip_bias_add else None
return output, output_bias
@@ -454,18 +424,14 @@ class MatmulAllreduceRowParallelOp(CustomRowParallelOp):
rank = torch.distributed.get_rank(group)
if torch.__version__ > "2.0":
global_rank = torch.distributed.get_global_rank(group, rank)
cls._HCOMM_INFO = group._get_backend(
torch.device("npu")).get_hccl_comm_name(global_rank)
cls._HCOMM_INFO = group._get_backend(torch.device("npu")).get_hccl_comm_name(global_rank)
else:
cls._HCOMM_INFO = group.get_hccl_comm_name(rank)
return cls._HCOMM_INFO
class SequenceColumnParallelOp(CustomColumnParallelOp):
def apply_impl(
self, input_: torch.Tensor
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
def apply_impl(self, input_: torch.Tensor) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
"""Linear layer with column parallelism.
Implemented multiple optimization projects for dense models, such as FlashComm and
@@ -490,13 +456,10 @@ class SequenceColumnParallelOp(CustomColumnParallelOp):
class Flashcomm2OshardQKVParallelOp(CustomColumnParallelOp):
def __init__(self, layer):
super().__init__(layer)
def apply_impl(
self, input_: torch.Tensor
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
def apply_impl(self, input_: torch.Tensor) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
"""Column-parallel linear with FlashComm2 OShard optimization."""
bias = self.bias if not self.skip_bias_add else None
@@ -505,12 +468,10 @@ class Flashcomm2OshardQKVParallelOp(CustomColumnParallelOp):
assert self.quant_method is not None
if enable_sp():
input_ = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
input_, True)
input_ = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(input_, True)
# Trigger async broadcast before matmul to overlap communication.
flashcomm2_oshard_manager.trigger_broadcast_for_layer(
self.layer.prefix)
flashcomm2_oshard_manager.trigger_broadcast_for_layer(self.layer.prefix)
output_parallel = self.quant_method.apply(self.layer, input_, bias)
if self.gather_output and self.tp_size > 1:
@@ -523,14 +484,11 @@ class Flashcomm2OshardQKVParallelOp(CustomColumnParallelOp):
class SequenceRowParallelOp(CustomRowParallelOp):
def __init__(self, layer):
super().__init__(layer)
self.unique_prefix = None
def apply_impl(
self, input_: torch.Tensor
) -> Union[torch.Tensor, tuple[torch.Tensor, Optional[Parameter]]]:
def apply_impl(self, input_: torch.Tensor) -> torch.Tensor | tuple[torch.Tensor, Parameter | None]:
"""Linear layer with column parallelism.
Implemented multiple optimization projects for dense models, such as FlashComm and
@@ -540,26 +498,21 @@ class SequenceRowParallelOp(CustomRowParallelOp):
if self.input_is_parallel:
input_parallel = input_
else:
splitted_input = split_tensor_along_last_dim(
input_, num_partitions=self.tp_size)
splitted_input = split_tensor_along_last_dim(input_, num_partitions=self.tp_size)
input_parallel = splitted_input[self.tp_rank].contiguous()
assert self.quant_method is not None
bias_ = None if (self.tp_rank > 0 or self.skip_bias_add) else self.bias
if self.tp_size == 1 or not self.reduce_results:
output = self.quant_method.apply(self.layer,
input_parallel,
bias=bias_)
output = self.quant_method.apply(self.layer, input_parallel, bias=bias_)
else:
output = torch.ops.vllm.matmul_and_reduce(input_parallel,
self.unique_prefix)
output = torch.ops.vllm.matmul_and_reduce(input_parallel, self.unique_prefix)
output_bias = self.bias if self.skip_bias_add else None
return output, output_bias
def matmul_and_reduce(self, input_parallel: torch.Tensor,
bias_: Optional[Parameter]) -> torch.Tensor:
def matmul_and_reduce(self, input_parallel: torch.Tensor, bias_: Parameter | None) -> torch.Tensor:
assert self.quant_method is not None
try:
forward_context = get_forward_context()
@@ -572,29 +525,24 @@ class SequenceRowParallelOp(CustomRowParallelOp):
x = input_parallel
if not sp_enabled:
output_parallel = self.layer.quant_method.apply(self.layer,
x,
bias=bias_)
output_parallel = self.layer.quant_method.apply(self.layer, x, bias=bias_)
return tensor_model_parallel_all_reduce(output_parallel)
pad_size = forward_context.pad_size
if pad_size > 0 and not (enable_dsa_cp()
and "o_proj" in self.layer.prefix):
if pad_size > 0 and not (enable_dsa_cp() and "o_proj" in self.layer.prefix):
x = F.pad(x, (0, 0, 0, pad_size))
world_size = self.layer.tp_size
comm_mode = "aiv"
hcom_name = get_tp_group().device_group._get_backend(
torch.device('npu')).get_hccl_comm_name(self.layer.tp_rank)
hcom_name = get_tp_group().device_group._get_backend(torch.device("npu")).get_hccl_comm_name(self.layer.tp_rank)
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
from vllm_ascend.quantization.methods import AscendW8A8LinearMethod
from vllm_ascend.quantization.method_adapters import AscendLinearMethod
from vllm_ascend.quantization.methods import AscendW8A8LinearMethod
# For unquant
if mmrs_fusion and isinstance(self.layer.quant_method,
UnquantizedLinearMethod):
if mmrs_fusion and isinstance(self.layer.quant_method, UnquantizedLinearMethod):
output = torch_npu.npu_mm_reduce_scatter_base(
x,
self.layer.weight.t(),
@@ -603,19 +551,22 @@ class SequenceRowParallelOp(CustomRowParallelOp):
reduce_op="sum",
bias=None,
comm_turn=0,
comm_mode=comm_mode)
comm_mode=comm_mode,
)
if bias_ is not None:
output.add_(bias_)
# For w8a8 quant
elif mmrs_fusion and (
isinstance(self.layer.quant_method, AscendLinearMethod)
and isinstance(self.layer.quant_method.quant_method,
AscendW8A8LinearMethod)):
isinstance(self.layer.quant_method, AscendLinearMethod)
and isinstance(self.layer.quant_method.quant_method, AscendW8A8LinearMethod)
):
if x.dtype != torch.int8:
x_quant = torch.ops.vllm.quantize(
x, self.layer.aclnn_input_scale,
x,
self.layer.aclnn_input_scale,
self.layer.aclnn_input_scale_reciprocal,
self.layer.aclnn_input_offset)
self.layer.aclnn_input_offset,
)
else:
x_quant = x
quant_bias = self.layer.quant_bias
@@ -631,14 +582,11 @@ class SequenceRowParallelOp(CustomRowParallelOp):
comm_turn=0,
x2_scale=deq_scale,
output_dtype=output_dtype,
comm_mode=comm_mode)
output = torch.add(
output,
torch.mul(quant_bias, deq_scale).to(self.layer.params_dtype))
comm_mode=comm_mode,
)
output = torch.add(output, torch.mul(quant_bias, deq_scale).to(self.layer.params_dtype))
else:
output_parallel = self.layer.quant_method.apply(self.layer,
x,
bias=bias_)
output_parallel = self.layer.quant_method.apply(self.layer, x, bias=bias_)
output = tensor_model_parallel_reduce_scatter(output_parallel, 0)
return output
@@ -651,13 +599,10 @@ class SequenceRowParallelOp(CustomRowParallelOp):
class ShardedCPRowParallelOp(CustomRowParallelOp):
@property
def comm_group(self):
# fake comm group to bypass tp logic
return SimpleNamespace(world_size=1,
rank_in_group=0,
device_group=None)
return SimpleNamespace(world_size=1, rank_in_group=0, device_group=None)
def apply_impl(
self,
@@ -677,13 +622,10 @@ class ShardedCPRowParallelOp(CustomRowParallelOp):
class ShardedCPColumnParallelOp(CustomColumnParallelOp):
@property
def comm_group(self):
# fake comm group to bypass tp logic
return SimpleNamespace(world_size=1,
rank_in_group=0,
device_group=None)
return SimpleNamespace(world_size=1, rank_in_group=0, device_group=None)
def apply_impl(
self,
@@ -700,12 +642,10 @@ class ShardedCPColumnParallelOp(CustomColumnParallelOp):
def _get_column_parallel_op(
prefix, layer
) -> Optional[Union[MLPColumnParallelOp, SequenceColumnParallelOp,
ShardedCPColumnParallelOp, Flashcomm2OshardQKVParallelOp]]:
) -> MLPColumnParallelOp | SequenceColumnParallelOp | ShardedCPColumnParallelOp | Flashcomm2OshardQKVParallelOp | None:
if enable_dsa_cp() and ("q_b_proj" in prefix or "kv_b_proj" in prefix):
return ShardedCPColumnParallelOp(layer)
if "gate_up_proj" in prefix and mlp_tp_enable(
) and not is_moe_layer(prefix):
if "gate_up_proj" in prefix and mlp_tp_enable() and not is_moe_layer(prefix):
return MLPColumnParallelOp(layer)
if flashcomm2_oshard_manager.flashcomm2_oshard_enable():
if any(p in prefix for p in ("qkv_proj", "conv1d", "query_key_value")):
@@ -714,7 +654,7 @@ def _get_column_parallel_op(
if "shared_expert" in prefix:
return None
sp_column_prefix = [
"gate_up_proj", # first MLP of most LLMs
"gate_up_proj", # first MLP of most LLMs
"in_proj", # gated deltanet of Qwen3 Next
"qkv_proj", # qkv linear of most LLMs
"conv1d", # gated deltanet of Qwen3 Next
@@ -729,9 +669,15 @@ def _get_column_parallel_op(
def _get_row_parallel_op(
prefix, layer
) -> Optional[Union[MLPRowParallelOp, OProjRowParallelOp,
Flashcomm2OProjRowParallelOp, MatmulAllreduceRowParallelOp,
SequenceRowParallelOp, ShardedCPRowParallelOp]]:
) -> (
MLPRowParallelOp
| OProjRowParallelOp
| Flashcomm2OProjRowParallelOp
| MatmulAllreduceRowParallelOp
| SequenceRowParallelOp
| ShardedCPRowParallelOp
| None
):
if enable_dsa_cp_with_layer_shard() and "o_proj" in prefix:
return ShardedCPRowParallelOp(layer)
if "down_proj" in prefix and mlp_tp_enable() and not is_moe_layer(prefix):
@@ -760,16 +706,21 @@ def _get_row_parallel_op(
def get_parallel_op(disable_tp, prefix, layer, direct):
if disable_tp or ("shared_experts" in prefix
and shared_expert_dp_enabled()):
if disable_tp or ("shared_experts" in prefix and shared_expert_dp_enabled()):
return None, 0, 1
custom_op: Optional[Union[MLPColumnParallelOp, SequenceColumnParallelOp,
MLPRowParallelOp, OProjRowParallelOp,
Flashcomm2OProjRowParallelOp,
Flashcomm2OshardQKVParallelOp,
MatmulAllreduceRowParallelOp,
SequenceRowParallelOp, ShardedCPRowParallelOp,
ShardedCPColumnParallelOp]] = None
custom_op: (
MLPColumnParallelOp
| SequenceColumnParallelOp
| MLPRowParallelOp
| OProjRowParallelOp
| Flashcomm2OProjRowParallelOp
| Flashcomm2OshardQKVParallelOp
| MatmulAllreduceRowParallelOp
| SequenceRowParallelOp
| ShardedCPRowParallelOp
| ShardedCPColumnParallelOp
| None
) = None
if direct == "row":
custom_op = _get_row_parallel_op(prefix, layer)
@@ -782,8 +733,7 @@ def get_parallel_op(disable_tp, prefix, layer, direct):
return None, get_tp_group().rank_in_group, get_tp_group().world_size
def get_replicated_op(disable_tp, prefix,
layer) -> Optional[Union[CustomReplicatedOp]]:
def get_replicated_op(disable_tp, prefix, layer) -> CustomReplicatedOp | None:
if disable_tp:
return None
@@ -791,24 +741,22 @@ def get_replicated_op(disable_tp, prefix,
def is_moe_layer(prefix: str) -> bool:
@lru_cache(maxsize=1)
def get_moe_params():
from vllm.config import get_current_vllm_config
vllm_config = get_current_vllm_config()
config = vllm_config.model_config.hf_text_config
n_routed_experts = getattr(config, 'n_routed_experts', 0)
first_k_dense_replace = getattr(config, 'first_k_dense_replace',
float('inf'))
moe_layer_freq = getattr(config, 'moe_layer_freq', 1)
n_routed_experts = getattr(config, "n_routed_experts", 0)
first_k_dense_replace = getattr(config, "first_k_dense_replace", float("inf"))
moe_layer_freq = getattr(config, "moe_layer_freq", 1)
return n_routed_experts, first_k_dense_replace, moe_layer_freq
match = re.search(r'layers\.(\d+)\.', prefix)
match = re.search(r"layers\.(\d+)\.", prefix)
if match is None:
return False
layer_idx = int(match.group(1))
n_routed_experts, first_k_dense_replace, moe_layer_freq = get_moe_params()
return (n_routed_experts is not None and layer_idx >= first_k_dense_replace
and layer_idx % moe_layer_freq == 0)
return n_routed_experts is not None and layer_idx >= first_k_dense_replace and layer_idx % moe_layer_freq == 0

302
vllm_ascend/ops/rotary_embedding.py
View File

@@ -17,13 +17,15 @@
import math
import os
from typing import Optional, Tuple
import torch
import torch_npu
from vllm.model_executor.layers.rotary_embedding import (
DeepseekScalingRotaryEmbedding, MRotaryEmbedding, RotaryEmbedding,
YaRNScalingRotaryEmbedding)
DeepseekScalingRotaryEmbedding,
MRotaryEmbedding,
RotaryEmbedding,
YaRNScalingRotaryEmbedding,
)
from vllm.model_executor.layers.rotary_embedding.common import ApplyRotaryEmb
from vllm.triton_utils import HAS_TRITON
@@ -31,8 +33,7 @@ if HAS_TRITON:
from vllm.model_executor.layers.rotary_embedding.mrope import triton_mrope
from vllm_ascend.platform import NPUPlatform
from vllm_ascend.utils import (AscendDeviceType, enable_custom_op,
get_ascend_device_type, has_rope, is_vl_model)
from vllm_ascend.utils import AscendDeviceType, enable_custom_op, get_ascend_device_type, has_rope, is_vl_model
# Currently, rope ops used on npu requires detached cos && sin as inputs.
# However, RotaryEmbedding in vllm use cos_sin_cache as a whole variable.
@@ -54,17 +55,13 @@ _cos_slice: torch.Tensor = None
_sin_slice: torch.Tensor = None
def set_cos_and_sin(vllm_config, max_num_reqs, decode_token_per_req, dtype,
device):
def set_cos_and_sin(vllm_config, max_num_reqs, decode_token_per_req, dtype, device):
global _cos_mla
global _sin_mla
global _cos
global _sin
if _cos_mla is not None or \
_sin_mla is not None or \
_cos is not None or \
_sin is not None:
if _cos_mla is not None or _sin_mla is not None or _cos is not None or _sin is not None:
return
model_config = vllm_config.model_config
@@ -72,36 +69,15 @@ def set_cos_and_sin(vllm_config, max_num_reqs, decode_token_per_req, dtype,
if model_config.use_mla:
rope_dim = model_config.hf_text_config.qk_rope_head_dim
_cos_mla = torch.ones(max_num_batched_tokens,
1,
1,
rope_dim,
dtype=dtype,
device=device)
_sin_mla = torch.zeros(max_num_batched_tokens,
1,
1,
rope_dim,
dtype=dtype,
device=device)
_cos_mla = torch.ones(max_num_batched_tokens, 1, 1, rope_dim, dtype=dtype, device=device)
_sin_mla = torch.zeros(max_num_batched_tokens, 1, 1, rope_dim, dtype=dtype, device=device)
elif not is_vl_model(vllm_config) and has_rope(vllm_config):
rope_dim = model_config.get_head_size()
# For models using partial rope like Qwen3-Next.
if hasattr(model_config.hf_text_config, "partial_rotary_factor"):
rope_dim = int(rope_dim *
model_config.hf_text_config.partial_rotary_factor)
_cos = torch.ones(1,
max_num_batched_tokens,
1,
rope_dim,
dtype=dtype,
device=device)
_sin = torch.zeros(1,
max_num_batched_tokens,
1,
rope_dim,
dtype=dtype,
device=device)
rope_dim = int(rope_dim * model_config.hf_text_config.partial_rotary_factor)
_cos = torch.ones(1, max_num_batched_tokens, 1, rope_dim, dtype=dtype, device=device)
_sin = torch.zeros(1, max_num_batched_tokens, 1, rope_dim, dtype=dtype, device=device)
def get_cos_and_sin_mla(positions, use_cache=False):
@@ -139,8 +115,7 @@ def _record_cos_and_sin_cache_interleaved(cos_sin_cache):
if _cos_cache is not None or _sin_cache is not None:
return
hidden_dim = cos_sin_cache.shape[-1] // 2
cos_cache, sin_cache = cos_sin_cache.view(-1, 2, hidden_dim).repeat(
1, 1, 2).chunk(2, dim=1)
cos_cache, sin_cache = cos_sin_cache.view(-1, 2, hidden_dim).repeat(1, 1, 2).chunk(2, dim=1)
_cos_cache = cos_cache.squeeze(1)
_sin_cache = sin_cache.squeeze(1)
@@ -151,16 +126,16 @@ def update_cos_sin(positions):
global _cos_slice
global _sin_slice
if _cos_sin_cache is None or \
_cos is None or \
_sin is None:
if _cos_sin_cache is None or _cos is None or _sin is None:
return
num_tokens = positions.size(0)
_cos[:, :num_tokens] = _cos_sin_cache.index_select(0, positions).view(
num_tokens, 2, -1).repeat(1, 1, 2).chunk(2, dim=-2)[0]
_sin[:, :num_tokens] = _cos_sin_cache.index_select(0, positions).view(
num_tokens, 2, -1).repeat(1, 1, 2).chunk(2, dim=-2)[1]
_cos[:, :num_tokens] = (
_cos_sin_cache.index_select(0, positions).view(num_tokens, 2, -1).repeat(1, 1, 2).chunk(2, dim=-2)[0]
)
_sin[:, :num_tokens] = (
_cos_sin_cache.index_select(0, positions).view(num_tokens, 2, -1).repeat(1, 1, 2).chunk(2, dim=-2)[1]
)
_cos_slice = _cos[:, :num_tokens]
_sin_slice = _sin[:, :num_tokens]
@@ -170,8 +145,7 @@ def get_cos_and_sin_slice():
def _custom_rotary_embedding_enabled(query, neox_style, head_size):
return query.dtype == torch.float16 and neox_style and head_size % 32 == 0 and enable_custom_op(
)
return query.dtype == torch.float16 and neox_style and head_size % 32 == 0 and enable_custom_op()
def _rope_forward_oot(
@@ -180,8 +154,8 @@ def _rope_forward_oot(
query: torch.Tensor,
key: torch.Tensor,
is_neox_style: bool,
offsets: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
offsets: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
query_shape, key_shape = query.shape, key.shape
if self.cos_sin_cache.device != query.device:
self.cos_sin_cache = self.cos_sin_cache.to(query.device)
@@ -189,8 +163,7 @@ def _rope_forward_oot(
self.cos_sin_cache = self.cos_sin_cache.to(query.dtype)
cos, sin = get_cos_and_sin_slice()
# adopt custom kernel path for rotary_embedding
if _custom_rotary_embedding_enabled(
query, is_neox_style, self.head_size):
if _custom_rotary_embedding_enabled(query, is_neox_style, self.head_size):
query, key = torch.ops._C_ascend.rotary_embedding(
positions,
query,
@@ -201,43 +174,40 @@ def _rope_forward_oot(
)
return query.view(query_shape), key.view(key_shape)
if offsets is not None:
raise NotImplementedError(
"Batched rotary embedding is currently not supported on NPU.")
raise NotImplementedError("Batched rotary embedding is currently not supported on NPU.")
else:
if is_neox_style and self.head_size == 128 and self.cos_sin_cache.shape[
-1] == 128 and cos is not None and sin is not None:
if (
is_neox_style
and self.head_size == 128
and self.cos_sin_cache.shape[-1] == 128
and cos is not None
and sin is not None
):
# If cos and sin are generated outside, use npu_apply_rotary_pos_emb to avoid redundant calculation.
# This method requires head_size and rotary_dim equal 128 and neox_style is True
query = query.contiguous().view(1, query.shape[0], -1,
self.head_size)
query = query.contiguous().view(1, query.shape[0], -1, self.head_size)
key = key.contiguous().view(1, key.shape[0], -1, self.head_size)
# Although this function modifies in-place, please retain the function's return value.
# Otherwise, the graph fusion operation may fail.
query, key = torch_npu.npu_apply_rotary_pos_emb(
query, key, cos, sin)
query, key = torch_npu.npu_apply_rotary_pos_emb(query, key, cos, sin)
elif self.rotary_dim < self.head_size:
if HAS_TRITON:
if HAS_TRITON:
cos = cos.view(-1, self.rotary_dim)
sin = sin.view(-1, self.rotary_dim)
q = query.contiguous().view(query.shape[0], -1,
self.head_size)
q = query.contiguous().view(query.shape[0], -1, self.head_size)
k = key.contiguous().view(key.shape[0], -1, self.head_size)
query, key = torch.ops.vllm.rope_forward_triton(q,
k,
cos,
sin,
rope_dim=self.rotary_dim,
is_neox_style=True)
query, key = torch.ops.vllm.rope_forward_triton(
q, k, cos, sin, rope_dim=self.rotary_dim, is_neox_style=True
)
return query.view(query_shape), key.view(key_shape)
else:
num_tokens = query.shape[0]
query = query.view(num_tokens, -1, self.head_size)
key = key.view(num_tokens, -1, self.head_size)
q_rot = query[..., :self.rotary_dim]
q_pass = query[..., self.rotary_dim:]
k_rot = key[..., :self.rotary_dim]
k_pass = key[..., self.rotary_dim:]
q_rot = query[..., : self.rotary_dim]
q_pass = query[..., self.rotary_dim :]
k_rot = key[..., : self.rotary_dim]
k_pass = key[..., self.rotary_dim :]
q_rot = q_rot.contiguous().view(num_tokens, -1)
k_rot = k_rot.contiguous().view(num_tokens, -1)
# only the rotary part is processed here,
@@ -271,7 +241,6 @@ def _rope_forward_oot(
class AscendRotaryEmbedding(RotaryEmbedding):
def __init__(
self,
head_size: int,
@@ -281,8 +250,7 @@ class AscendRotaryEmbedding(RotaryEmbedding):
is_neox_style: bool,
dtype: torch.dtype,
) -> None:
super().__init__(head_size, rotary_dim, max_position_embeddings, base,
is_neox_style, dtype)
super().__init__(head_size, rotary_dim, max_position_embeddings, base, is_neox_style, dtype)
_record_cos_sin_cache(self.cos_sin_cache)
_record_cos_and_sin_cache_interleaved(self.cos_sin_cache)
@@ -291,18 +259,16 @@ class AscendRotaryEmbedding(RotaryEmbedding):
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor,
offsets: Optional[torch.Tensor] = None,
is_neox_style_override: Optional[bool] = None,
offsets: torch.Tensor | None = None,
is_neox_style_override: bool | None = None,
):
is_neox_style = self.is_neox_style
if is_neox_style_override is not None:
is_neox_style = is_neox_style_override
return _rope_forward_oot(self, positions, query, key, is_neox_style,
offsets)
return _rope_forward_oot(self, positions, query, key, is_neox_style, offsets)
class AscendYaRNRotaryEmbedding(YaRNScalingRotaryEmbedding):
def __init__(
self,
head_size: int,
@@ -322,10 +288,11 @@ class AscendYaRNRotaryEmbedding(YaRNScalingRotaryEmbedding):
"extrapolation_factor": extrapolation_factor,
"attn_factor": attn_factor,
"beta_fast": beta_fast,
"beta_slow": beta_slow
"beta_slow": beta_slow,
}
super().__init__(head_size, rotary_dim, max_position_embeddings, base,
is_neox_style, scaling_factor, dtype, **extra_kwargs)
super().__init__(
head_size, rotary_dim, max_position_embeddings, base, is_neox_style, scaling_factor, dtype, **extra_kwargs
)
_record_cos_sin_cache(self.cos_sin_cache)
def forward_oot(
@@ -333,16 +300,13 @@ class AscendYaRNRotaryEmbedding(YaRNScalingRotaryEmbedding):
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor,
offsets: Optional[torch.Tensor] = None,
is_neox_style_override: Optional[bool] = None,
offsets: torch.Tensor | None = None,
is_neox_style_override: bool | None = None,
):
return AscendRotaryEmbedding.forward_oot(self, positions, query, key,
offsets,
is_neox_style_override)
return AscendRotaryEmbedding.forward_oot(self, positions, query, key, offsets, is_neox_style_override)
class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
def __init__(
self,
head_size: int,
@@ -370,18 +334,17 @@ class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
self.beta_slow = beta_slow
# Get n-d magnitude scaling corrected for interpolation.
self.mscale = float(
self._yarn_get_mscale(self.scaling_factor, float(mscale)) /
self._yarn_get_mscale(self.scaling_factor, float(mscale_all_dim)) *
attn_factor)
super(DeepseekScalingRotaryEmbedding,
self).__init__(head_size, rotary_dim, max_position_embeddings,
base, is_neox_style, dtype)
self._yarn_get_mscale(self.scaling_factor, float(mscale))
/ self._yarn_get_mscale(self.scaling_factor, float(mscale_all_dim))
* attn_factor
)
super(DeepseekScalingRotaryEmbedding, self).__init__(
head_size, rotary_dim, max_position_embeddings, base, is_neox_style, dtype
)
# NOTE: For ascend friendly computing, reorder sin and cos cache
self.max_seq_len = math.ceil(max_position_embeddings * scaling_factor)
self._set_cos_sin_cache(self.max_seq_len,
device=NPUPlatform.device_type,
dtype=dtype)
self._set_cos_sin_cache(self.max_seq_len, device=NPUPlatform.device_type, dtype=dtype)
def _yarn_get_mscale(self, scale: float = 1, mscale: float = 1) -> float:
if scale <= 1:
@@ -390,56 +353,35 @@ class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
def _rotate_half(self, x):
"""Rotates half the hidden dims of the input."""
x1 = x[..., :x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2:]
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
def _yarn_linear_ramp_mask(self, min_value, max_value, dim):
# Note: The if conditional branch is not used here
# to solve MTP compilation error.
max_value += (min_value == max_value).float() * 0.001
linear_func = (torch.arange(dim, dtype=torch.float32) -
min_value) / (max_value - min_value)
linear_func = (torch.arange(dim, dtype=torch.float32) - min_value) / (max_value - min_value)
ramp_func = torch.clamp(linear_func, 0, 1)
return ramp_func
# Inverse dim formula to find dim based on number of rotations
def _yarn_find_correction_dim(self,
num_rotations,
dim,
base=10000,
max_position_embeddings=2048):
def _yarn_find_correction_dim(self, num_rotations, dim, base=10000, max_position_embeddings=2048):
# Note: use torch instead of math to solve MTP compilation error.
return (dim * torch.log(
torch.tensor(max_position_embeddings) /
(num_rotations * 2 * torch.pi))) / (2 *
torch.log(torch.tensor(base)))
return (dim * torch.log(torch.tensor(max_position_embeddings) / (num_rotations * 2 * torch.pi))) / (
2 * torch.log(torch.tensor(base))
)
# Find dim range bounds based on rotations
def _yarn_find_correction_range(self,
low_rot,
high_rot,
dim,
base=10000,
max_position_embeddings=2048):
def _yarn_find_correction_range(self, low_rot, high_rot, dim, base=10000, max_position_embeddings=2048):
# Note: use torch instead of math to solve MTP compilation error.
low = torch.floor(
self._yarn_find_correction_dim(low_rot, dim, base,
max_position_embeddings))
high = torch.ceil(
self._yarn_find_correction_dim(high_rot, dim, base,
max_position_embeddings))
low = torch.floor(self._yarn_find_correction_dim(low_rot, dim, base, max_position_embeddings))
high = torch.ceil(self._yarn_find_correction_dim(high_rot, dim, base, max_position_embeddings))
# Note: use torch instead of max/min to solve MTP compilation error.
return torch.clamp(low, min=0), torch.clamp(high, max=dim - 1)
# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def _apply_rotary_pos_emb(self,
q,
k,
cos,
sin,
position_ids,
unsqueeze_dim=1):
def _apply_rotary_pos_emb(self, q, k, cos, sin, position_ids, unsqueeze_dim=1):
"""Applies Rotary Position Embedding to the query and key tensors.
Args:
q (`torch.Tensor`): The query tensor.
@@ -451,11 +393,11 @@ class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
used to pass offsetted position ids when working with a KV-cache.
unsqueeze_dim (`int`, *optional*, defaults to 1):
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example,
note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim].
Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1
makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly,
if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
Returns:
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
"""
@@ -488,10 +430,10 @@ class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
def _set_cos_sin_cache(self, max_seq_len, device, dtype):
dim = self.rotary_dim
freq_extra = 1.0 / (self.base**(
torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
freq_inter = 1.0 / (self.scaling_factor * self.base**(
torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
freq_extra = 1.0 / (self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
freq_inter = 1.0 / (
self.scaling_factor * self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim)
)
low, high = self._yarn_find_correction_range(
self.beta_fast,
@@ -500,10 +442,8 @@ class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
self.base,
self.max_position_embeddings,
)
inv_freq_mask = 1.0 - self._yarn_linear_ramp_mask(
low, high, dim // 2).to(device=device, dtype=torch.float32)
inv_freq = freq_inter * (1 -
inv_freq_mask) + freq_extra * inv_freq_mask
inv_freq_mask = 1.0 - self._yarn_linear_ramp_mask(low, high, dim // 2).to(device=device, dtype=torch.float32)
inv_freq = freq_inter * (1 - inv_freq_mask) + freq_extra * inv_freq_mask
self.register_buffer("inv_freq", inv_freq, persistent=False)
t = torch.arange(max_seq_len, device=device, dtype=torch.float32)
@@ -513,20 +453,16 @@ class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
sin_cached = torch.cat([freqs, freqs], dim=-1).sin() * self.mscale
cos_cached = cos_cached.to(dtype)
sin_cached = sin_cached.to(dtype)
cache = torch.cat(
[freqs.cos() * self.mscale,
freqs.sin() * self.mscale], dim=-1).to(dtype)
cache = torch.cat([freqs.cos() * self.mscale, freqs.sin() * self.mscale], dim=-1).to(dtype)
self.register_buffer("cos_sin_cache", cache, persistent=False)
self.register_buffer("cos_cached", cos_cached, persistent=False)
self.register_buffer("sin_cached", sin_cached, persistent=False)
_record_cos_sin_cache(cache)
_record_cos_and_sin_cache(cos_cached, sin_cached)
def forward(self,
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor,
offsets: Optional[torch.Tensor] = None):
def forward(
self, positions: torch.Tensor, query: torch.Tensor, key: torch.Tensor, offsets: torch.Tensor | None = None
):
if len(key.shape) == 2:
key = key[:, None, :]
# Note: we implement the non neox_style method with shuffle the last dim and neox style
@@ -535,26 +471,24 @@ class AscendDeepseekScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding):
is_neox_style = True
if self.is_neox_style is False:
b, h_q, d = query.shape
query = query.view(b, h_q, d // 2,
2).transpose(3, 2).reshape(b, h_q, d)
query = query.view(b, h_q, d // 2, 2).transpose(3, 2).reshape(b, h_q, d)
b, h_k, d = key.shape
key = key.view(b, h_k, d // 2, 2).transpose(3,
2).reshape(b, h_k, d)
q_pe, k_pe = _rope_forward_oot(self, positions, query, key,
is_neox_style, offsets)
key = key.view(b, h_k, d // 2, 2).transpose(3, 2).reshape(b, h_k, d)
q_pe, k_pe = _rope_forward_oot(self, positions, query, key, is_neox_style, offsets)
return q_pe, k_pe
class AscendMRotaryEmbedding(MRotaryEmbedding):
# Empirical safety threshold for large Triton grids on Ascend NPU
_ASCEND_TRITON_GRID_LIMIT = 65535
def forward_triton(self,
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor | None = None,
offsets: torch.Tensor | None = None):
def forward_triton(
self,
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor | None = None,
offsets: torch.Tensor | None = None,
):
assert positions.ndim == 2
assert key is not None
@@ -571,10 +505,9 @@ class AscendMRotaryEmbedding(MRotaryEmbedding):
assert self.mrope_section
# When the grid becomes large, enable TRITON_ALL_BLOCKS_PARALLEL
# When the grid becomes large, enable TRITON_ALL_BLOCKS_PARALLEL
# to avoid scheduler/runtime failures.
if (query_shape[0] > self._ASCEND_TRITON_GRID_LIMIT and
os.environ.get("TRITON_ALL_BLOCKS_PARALLEL") != "1"):
if query_shape[0] > self._ASCEND_TRITON_GRID_LIMIT and os.environ.get("TRITON_ALL_BLOCKS_PARALLEL") != "1":
os.environ["TRITON_ALL_BLOCKS_PARALLEL"] = "1"
q, k = triton_mrope(
@@ -600,35 +533,37 @@ class AscendMRotaryEmbedding(MRotaryEmbedding):
# todo: need cann update in 8.5.0
return self.forward_triton(positions, query, key)
if self.mrope_section != [16, 24, 24] or \
get_ascend_device_type() == AscendDeviceType.A5:
if self.mrope_section != [16, 24, 24] or get_ascend_device_type() == AscendDeviceType.A5:
return super().forward_oot(positions, query, key)
import torch_npu
mrope_section = [0, 0, 0
] if positions.ndim == 1 else self.mrope_section
mrope_section = [0, 0, 0] if positions.ndim == 1 else self.mrope_section
if self.cos_sin_cache.device != query.device: # type: ignore
self.cos_sin_cache = self.cos_sin_cache.to( # type: ignore
query.device) # type: ignore
query.device
) # type: ignore
if self.cos_sin_cache.dtype != query.dtype: # type: ignore
self.cos_sin_cache = self.cos_sin_cache.to( # type: ignore
query.dtype) # type: ignore
query.dtype
) # type: ignore
query, key = torch_npu.npu_mrope(positions.contiguous(),
query.contiguous(),
key.contiguous(),
self.cos_sin_cache.contiguous(),
self.head_size,
mrope_section=mrope_section,
rotary_mode='half')
query, key = torch_npu.npu_mrope(
positions.contiguous(),
query.contiguous(),
key.contiguous(),
self.cos_sin_cache.contiguous(),
self.head_size,
mrope_section=mrope_section,
rotary_mode="half",
)
return query, key
class AscendApplyRotaryEmb(ApplyRotaryEmb):
def __init__(
self,
enforce_enable: bool = False,
@@ -647,8 +582,7 @@ class AscendApplyRotaryEmb(ApplyRotaryEmb):
cos: torch.Tensor,
sin: torch.Tensor,
) -> torch.Tensor:
x, cos, sin, origin_shape, origin_dtype = self._pre_process(
x, cos, sin)
x, cos, sin, origin_shape, origin_dtype = self._pre_process(x, cos, sin)
head_dim = x.shape[-1]
# cos, sin: [seq_len, head_dim // 2]