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 17, 08:00 UTC
Fix the `KeyError` when loading bloom-based models (#441)
Wen Sun committed
Jul 14, 2023 at 12:58 UTC
dbed69058c88ddf42914e6ab3a9b6ea12e15b12a
1 file changed
+10
-3
vllm/model_executor/models/bloom.py
+10
-3
@@ -284,10 +284,17 @@ class BloomForCausalLM(nn.Module):
284
state_dict = self.state_dict()
285
for name, loaded_weight in hf_model_weights_iterator(
286
model_name_or_path, cache_dir, use_np_cache):
287
- if not name.startswith("transformer."):
288
- name = "transformer." + name
287
+ if name == "lm_head.weight":
288
+ # Since hidden_states are parallelized, we need to
289
+ # load lm_head.weight in parallel.
290
+ self._column_parallel_weights.append(name)
291
+ # If lm_head is provided, use it instead.
292
+ param = self.lm_head_weight
293
+ else:
294
+ if not name.startswith("transformer."):
295
+ name = "transformer." + name
296
+ param = state_dict[name]
297
290
- param = state_dict[name]
298
if "query_key_value" in name:
299
# NOTE(woosuk): BLOOM's fused QKV has the shape of
300
# [num_heads * 3 * head_size, hidden_size], while the