xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-12 05:07:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Make the deviceprofiler include LoRA memory. (#14469)

Browse Source 
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li 2025-03-08 15:12:22 +08:00 committed by GitHub
parent c908a07f57
commit b8b0ccbd2d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 29 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
vllm/worker/model_runner.py
View File

@ -1111,41 +1111,41 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
with DeviceMemoryProfiler(self.device) as m: with DeviceMemoryProfiler(self.device) as m:
time_before_load = time.perf_counter() time_before_load = time.perf_counter()
self.model = get_model(vllm_config=self.vllm_config) self.model = get_model(vllm_config=self.vllm_config)
if self.lora_config:
assert supports_lora(
self.model
), f"{self.model.__class__.__name__} does not support LoRA yet."
if supports_multimodal(self.model):
logger.warning(
"Regarding multimodal models, vLLM currently "
"only supports adding LoRA to language model.")
# It's necessary to distinguish between the
# max_position_embeddings of VLMs and LLMs.
if hasattr(self.model.config, "max_position_embeddings"):
max_pos_embeddings = (
self.model.config.max_position_embeddings)
else:
max_pos_embeddings = (
self.model.config.text_config.max_position_embeddings)
self.lora_manager = LRUCacheWorkerLoRAManager(
self.scheduler_config.max_num_seqs,
self.scheduler_config.max_num_batched_tokens,
self.vocab_size,
self.lora_config,
self.device,
self.model.embedding_modules,
self.model.embedding_padding_modules,
max_position_embeddings=max_pos_embeddings,
)
self.model = self.lora_manager.create_lora_manager(self.model)
time_after_load = time.perf_counter() time_after_load = time.perf_counter()
self.model_memory_usage = m.consumed_memory self.model_memory_usage = m.consumed_memory
logger.info("Model loading took %.4f GB and %.6f seconds", logger.info("Model loading took %.4f GB and %.6f seconds",
self.model_memory_usage / float(2**30), self.model_memory_usage / float(2**30),
time_after_load - time_before_load) time_after_load - time_before_load)
if self.lora_config:
assert supports_lora(
self.model
), f"{self.model.__class__.__name__} does not support LoRA yet."
if supports_multimodal(self.model):
logger.warning("Regarding multimodal models, vLLM currently "
"only supports adding LoRA to language model.")
# It's necessary to distinguish between the max_position_embeddings
# of VLMs and LLMs.
if hasattr(self.model.config, "max_position_embeddings"):
max_pos_embeddings = self.model.config.max_position_embeddings
else:
max_pos_embeddings = (
self.model.config.text_config.max_position_embeddings)
self.lora_manager = LRUCacheWorkerLoRAManager(
self.scheduler_config.max_num_seqs,
self.scheduler_config.max_num_batched_tokens,
self.vocab_size,
self.lora_config,
self.device,
self.model.embedding_modules,
self.model.embedding_padding_modules,
max_position_embeddings=max_pos_embeddings,
)
self.model = self.lora_manager.create_lora_manager(self.model)
if self.prompt_adapter_config: if self.prompt_adapter_config:
self.prompt_adapter_manager = LRUCacheWorkerPromptAdapterManager( self.prompt_adapter_manager = LRUCacheWorkerPromptAdapterManager(
self.scheduler_config.max_num_seqs, self.scheduler_config.max_num_seqs,