A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 05:00 UTC
Remove dead bitsandbytes CxB code from 8-bit inference path (#34633)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Tim Dettmers committed
Feb 17, 2026 at 04:49 UTC
d44a5b6c474ae02f1daf61f6db623a4311294ef8
1 file changed
-10
vllm/model_executor/layers/quantization/bitsandbytes.py
-10
@@ -336,16 +336,6 @@ class BitsAndBytesLinearMethod(LinearMethodBase):
336
337
current_index += output_size
338
339
- # only update the matmul_states if it is not profile_run
340
- if (
341
- generation > 0
342
- and not self.quant_config.llm_int8_has_fp16_weight
343
- and matmul_states[i].CB is not None
344
- and matmul_states[i].CxB is not None
345
- ):
346
- del matmul_states[i].CB
347
- qweight[offsets[i] : offsets[i + 1]] = matmul_states[i].CxB
348
-
339
out = out.to(original_type)
340
341
if reshape_after_matmul: