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 9, 14:45 UTC
Remove unused logic in `models/mistral.py` (#33095)
Signed-off-by: Andy Lo <andy@mistral.ai>
Andy Lo committed
Jan 26, 2026 at 17:01 UTC
d56afd45fd4efee581129c401613be356b95350d
1 file changed
-8
vllm/model_executor/models/mistral.py
-8
@@ -156,16 +156,8 @@ class MistralDecoderLayer(LlamaDecoderLayer):
156
)
157
158
self.layer_idx = int(prefix.split(sep=".")[-1])
159
- quant_config = self.get_quant_config(vllm_config)
159
config = config or vllm_config.model_config.hf_config
160
162
- do_fusion = getattr(
163
- quant_config, "enable_quantization_scaling_fusion", False
164
- ) and vllm_config.cache_config.cache_dtype.startswith("fp8")
165
- if do_fusion:
166
- self.input_layernorm.quant_scaling_from = self.self_attn.qkv_proj
167
- self.post_attention_layernorm.quant_scaling_from = self.mlp.gate_up_proj
168
-
161
if getattr(config, "ada_rms_norm_t_cond", False):
162
self.ada_rms_norm_t_cond = nn.Sequential(
163
ColumnParallelLinear(