@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / ea928f608c
Fork 0

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 16, 08:15 UTC

[Bugfix] Gpt-j-6B patch kv_scale to k_scale path (#10063)

Signed-off-by: Alex Rakowski <alex.rakowski@amd.com> Signed-off-by: Alex Rakowski <182798202+arakowsk-amd@users.noreply.github.com>

arakowsk-amd committed Nov 5, 2024 at 21:10 UTC ea928f608c44b825d28609460e0d375a5f877940
1 file changed +5 -1
vllm/model_executor/models/gpt_j.py
+5 -1
@@ -36,7 +36,8 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
36 from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
37 from vllm.model_executor.layers.vocab_parallel_embedding import (
38 ParallelLMHead, VocabParallelEmbedding)
39 -from vllm.model_executor.model_loader.weight_utils import default_weight_loader
39 +from vllm.model_executor.model_loader.weight_utils import (
40 + default_weight_loader, maybe_remap_kv_scale_name)
41 from vllm.model_executor.sampling_metadata import SamplingMetadata
42 from vllm.sequence import IntermediateTensors
43
@@ -308,6 +309,9 @@ class GPTJForCausalLM(nn.Module, SupportsPP):
309 weight_loader(param, loaded_weight, shard_id)
310 break
311 else:
312 + name = maybe_remap_kv_scale_name(name, params_dict)
313 + if name is None:
314 + continue
315 # Skip loading extra bias for GPTQ models.
316 if name.endswith(".bias") and name not in params_dict:
317 continue