[ Misc ] Enable Quantizing All Layers of DeekSeekv2 (#6423)

Author: robertgshaw2-redhat

.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh

@@ -46,6 +46,6 @@ while getopts "m:b:l:f:t:" OPT; do
 done
 
 lm_eval --model vllm \
-  --model_args pretrained=$MODEL,tensor_parallel_size=$TP_SIZE,add_bos_token=true,distributed_executor_backend="ray",trust_remote_code=true \
+  --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 \
   --tasks gsm8k --num_fewshot $FEWSHOT --limit $LIMIT \
   --batch_size $BATCH_SIZE

vllm/model_executor/model_loader/weight_utils.py

@@ -431,6 +431,11 @@ def convert_pyslice_to_tensor(x: Any) -> torch.Tensor:
 def default_weight_loader(param: torch.Tensor,
                           loaded_weight: torch.Tensor) -> None:
     """Default weight loader."""
+    # If the weight on disk does not have a shape, give it one
+    # (such scales for AutoFp8).
+    if len(loaded_weight.shape) == 0:
+        loaded_weight = loaded_weight.reshape(1)
+
     assert param.size() == loaded_weight.size()
     param.data.copy_(loaded_weight)