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, 20:45 UTC
[ Misc ] Enable Quantizing All Layers of DeekSeekv2 (#6423)
Robert Shaw committed
Jul 14, 2024 at 17:38 UTC
73030b7dae676f730ea43652af056501584ecbfe
2 files changed
+6
-1
.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh
+1
-1
@@ -46,6 +46,6 @@ while getopts "m:b:l:f:t:" OPT; do
46
done
47
48
lm_eval --model vllm \
49
- --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,add_bos_token=true,distributed_executor_backend="ray",trust_remote_code=true \
49
+ --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,add_bos_token=true,distributed_executor_backend="ray",trust_remote_code=true,max_model_len=4096 \
50
--tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
51
--batch_size $BATCH_SIZE
vllm/model_executor/model_loader/weight_utils.py
+5
@@ -431,6 +431,11 @@ def convert_pyslice_to_tensor(x: Any) -> torch.Tensor:
431
def default_weight_loader(param: torch.Tensor,
432
loaded_weight: torch.Tensor) -> None:
433
"""Default weight loader."""
434
+ # If the weight on disk does not have a shape, give it one
435
+ # (such scales for AutoFp8).
436
+ if len(loaded_weight.shape) == 0:
437
+ loaded_weight = loaded_weight.reshape(1)
438
+
439
assert param.size() == loaded_weight.size()
440
param.data.copy_(loaded_weight)
441