[Bugfix] Propagate quant_config to LFM2 ShortConv projections (#48917)
Signed-off-by: Alex Yuan <alex.yuan@liquid.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Alex Yuan committed
Jul 20, 2026 at 23:51 UTC
8def3cdde2a0acf0019447d8fcfc57e4df6bcfc2
3 files changed
+6
vllm/model_executor/layers/mamba/short_conv.py
+4
@@ -23,6 +23,7 @@ from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
23
causal_conv1d_fn,
24
causal_conv1d_update,
25
)
26
+from vllm.model_executor.layers.quantization import QuantizationConfig
27
from vllm.platforms import current_platform
28
from vllm.utils.torch_utils import direct_register_custom_op
29
from vllm.v1.attention.backend import AttentionMetadata
@@ -42,6 +43,7 @@ class ShortConv(MambaBase, CustomOp):
43
layer_idx: int,
44
model_config: ModelConfig | None = None,
45
cache_config: CacheConfig | None = None,
46
+ quant_config: QuantizationConfig | None = None,
47
prefix: str = "",
48
):
49
super().__init__()
@@ -67,12 +69,14 @@ class ShortConv(MambaBase, CustomOp):
69
input_size=dim,
70
output_sizes=[dim] * 3,
71
bias=self.bias,
72
+ quant_config=quant_config,
73
prefix=f"{prefix}.in_proj",
74
)
75
self.out_proj = RowParallelLinear(
76
input_size=dim,
77
output_size=dim,
78
bias=self.bias,
79
+ quant_config=quant_config,
80
prefix=f"{prefix}.out_proj",
81
)
82
vllm/model_executor/models/lfm2.py
+1
@@ -257,6 +257,7 @@ class Lfm2ShortConvDecoderLayer(nn.Module):
257
layer_idx=layer_idx,
258
model_config=model_config,
259
cache_config=cache_config,
260
+ quant_config=quant_config,
261
prefix=f"{prefix}.conv",
262
)
263
vllm/model_executor/models/lfm2_moe.py
+1
@@ -351,6 +351,7 @@ class Lfm2MoeShortConvDecoderLayer(nn.Module):
351
layer_idx=layer_idx,
352
model_config=model_config,
353
cache_config=cache_config,
354
+ quant_config=quant_config,
355
prefix=f"{prefix}.conv",
356
)
357