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 15, 18:30 UTC
[Bugfix] Fix interface for Olmo2 on V1 (#14976)
Signed-off-by: Roger Wang <ywang@roblox.com>
Roger Wang committed
Mar 17, 2025 at 11:26 UTC
5340b0e2214ec71117e7f0a953cf3033e3194d2a
1 file changed
+9
-5
vllm/model_executor/models/olmo2.py
+9
-5
@@ -42,7 +42,7 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
42
RowParallelLinear)
43
from vllm.model_executor.layers.logits_processor import LogitsProcessor
44
from vllm.model_executor.layers.rotary_embedding import get_rope
45
-from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
45
+from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
46
from vllm.model_executor.layers.vocab_parallel_embedding import (
47
ParallelLMHead, VocabParallelEmbedding)
48
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -283,17 +283,19 @@ class Olmo2Model(nn.Module):
283
input_ids: torch.Tensor,
284
positions: torch.Tensor,
285
intermediate_tensors: Optional[IntermediateTensors],
286
+ inputs_embeds: Optional[torch.Tensor] = None,
287
) -> Union[torch.Tensor, IntermediateTensors]:
288
"""
289
:param input_ids: A tensor of shape `(batch_size, seq_len)`.
290
"""
291
if get_pp_group().is_first_rank:
292
+ if inputs_embeds is not None:
293
+ hidden_states = inputs_embeds
294
# Get embeddings of input.
295
# shape: (batch_size, seq_len, d_model)
293
- inputs_embeds = self.embed_tokens(input_ids)
296
+ else:
297
+ hidden_states = self.embed_tokens(input_ids)
298
295
- # embed positions
296
- hidden_states = inputs_embeds
299
else:
300
assert intermediate_tensors is not None
301
hidden_states = intermediate_tensors["hidden_states"]
@@ -337,7 +339,7 @@ class Olmo2ForCausalLM(nn.Module, SupportsPP):
339
prefix=maybe_prefix(prefix, "lm_head"),
340
)
341
self.logits_processor = LogitsProcessor(config.vocab_size)
340
- self.sampler = Sampler()
342
+ self.sampler = get_sampler()
343
self.make_empty_intermediate_tensors = (
344
self.model.make_empty_intermediate_tensors)
345
@@ -346,11 +348,13 @@ class Olmo2ForCausalLM(nn.Module, SupportsPP):
348
input_ids: torch.Tensor,
349
positions: torch.Tensor,
350
intermediate_tensors: Optional[IntermediateTensors] = None,
351
+ inputs_embeds: Optional[torch.Tensor] = None,
352
) -> Union[torch.Tensor, IntermediateTensors]:
353
hidden_states = self.model(
354
input_ids=input_ids,
355
positions=positions,
356
intermediate_tensors=intermediate_tensors,
357
+ inputs_embeds=inputs_embeds,
358
)
359
return hidden_states
360