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

[Bugfix] Remove ModelSlim-"M4 Quantization". (#4589)

Browse Source 
The M4 quantization method in ModelSlim adds bias to model weights that
originally do not have a linear bias. PR #4235 supported PD-MIX
quantization and M4 quantization, adding bias to `w8a8.py` and
`w8a8_dynamic.py`, and implementing adaptations in `ops/linear.py` to
prevent it from being reset to `None` by
`self.register_parameter("bias", None)`. However, this modification
introduced an issue where the bias was still being reset to `None` in
certain scenarios, causing errors during service startup. Therefore,
support for M4 quantization is temporarily being reverted in this PR.
___
- vLLM version: v0.11.2

Signed-off-by: SlightwindSec <slightwindsec@gmail.com>
This commit is contained in:
Slightwind
2025-12-01 23:45:02 +08:00
committed by GitHub
parent 8813832387
commit 12ca99c94e
3 changed files with 7 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm_ascend/ops/linear.py
View File

@@ -277,20 +277,18 @@ class AscendRowParallelLinear(RowParallelLinear):
weight_loader=(
self.weight_loader_v2 if self.quant_method.__class__.__name__
in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
bias_initialized_by_quant = ("bias" in self._parameters
and self._parameters["bias"] is not None)
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")
if bias and not bias_initialized_by_quant:
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,
})
elif not bias and not bias_initialized_by_quant:
else:
self.register_parameter("bias", None)
if self.custom_op is not None:
@@ -368,9 +366,7 @@ class AscendColumnParallelLinear(ColumnParallelLinear):
weight_loader=(
self.weight_loader_v2 if self.quant_method.__class__.__name__
in WEIGHT_LOADER_V2_SUPPORTED else self.weight_loader))
bias_initialized_by_quant = ("bias" in self._parameters
and self._parameters["bias"] is not None)
if bias and not bias_initialized_by_quant:
if bias:
self.bias = Parameter(
torch.empty(self.output_size_per_partition,
dtype=params_dtype))
@@ -378,7 +374,7 @@ class AscendColumnParallelLinear(ColumnParallelLinear):
"output_dim": 0,
"weight_loader": self.weight_loader,
})
elif not bias and not bias_initialized_by_quant:
else:
self.register_parameter("bias", None)
if self.custom_op is not None:
@@ -449,16 +445,14 @@ class AscendReplicatedLinear(ReplicatedLinear):
self.params_dtype,
weight_loader=self.weight_loader)
bias_initialized_by_quant = ("bias" in self._parameters
and self._parameters["bias"] is not None)
if bias and not bias_initialized_by_quant:
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,
})
elif not bias and not bias_initialized_by_quant:
else:
self.register_parameter("bias", None)
if self.custom_op is not None: