[Model Runner V2] Add LoRAState to consolidate lora logic (#33062)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
This commit is contained in:
Woosuk Kwon
2026-01-25 22:21:12 -08:00
committed by GitHub
parent 254db42ede
commit a9b53dd435
3 changed files with 53 additions and 42 deletions
+6 -2
View File
@@ -51,6 +51,7 @@ from vllm.v1.worker.gpu.kv_connector import (
KVConnector,
get_kv_connector,
)
from vllm.v1.worker.gpu.lora_utils import LoraState
from vllm.v1.worker.gpu.mm.encoder_runner import EncoderRunner
from vllm.v1.worker.gpu.mm.mrope_utils import MRopeState
from vllm.v1.worker.gpu.sample.output import SamplerOutput
@@ -168,6 +169,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
max_num_logits=self.max_num_reqs * (self.num_speculative_steps + 1),
vocab_size=self.vocab_size,
)
# LoRA-related workers.
self.lora_state = LoraState(max_num_reqs=self.max_num_reqs)
# Buffers for CPU-to-GPU copies.
self.tmp_idx_mapping = UvaBufferPool(self.max_num_reqs, torch.int32)
@@ -426,6 +429,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
if self.supports_mm_inputs:
self.encoder_runner.remove_request(req_id)
self.prompt_logprobs_worker.remove_request(req_id)
self.lora_state.remove_request(req_id)
def free_states(self, scheduler_output: SchedulerOutput) -> None:
if self.supports_mm_inputs:
@@ -444,7 +448,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
prompt_len=prompt_len,
prefill_token_ids=new_req_data.prefill_token_ids,
num_computed_tokens=new_req_data.num_computed_tokens,
lora_request=new_req_data.lora_request,
)
req_index = self.req_states.req_id_to_index[req_id]
@@ -469,6 +472,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
self.prompt_logprobs_worker.add_request(
req_id, req_index, new_req_data.sampling_params
)
self.lora_state.add_request(req_id, req_index, new_req_data.lora_request)
if scheduler_output.scheduled_new_reqs:
self.req_states.apply_staged_writes()
@@ -841,7 +845,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
)
if self.lora_config:
# Activate LoRA adapters.
lora_inputs = self.req_states.make_lora_inputs(
lora_inputs = self.lora_state.make_lora_inputs(
input_batch.req_ids,
input_batch.idx_mapping_np,
input_batch.num_scheduled_tokens,