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, 22:15 UTC
[ci] fix linter (#13701)
Signed-off-by: youkaichao <youkaichao@gmail.com>
youkaichao committed
Feb 22, 2025 at 20:28 UTC
2382ad29d1769f2b46d51fcc2cedbd6e00d4f180
5 files changed
+11
-7
examples/offline_inference/data_parallel.py
+5
-4
@@ -48,15 +48,16 @@ def main(dp_size, dp_rank, dp_master_ip, dp_master_port, GPUs_per_dp_rank):
48
max_tokens=16 * (dp_rank + 1))
49
50
# Create an LLM.
51
- llm = LLM(model="facebook/opt-125m", tensor_parallel_size=2, enforce_eager=True)
51
+ llm = LLM(model="facebook/opt-125m",
52
+ tensor_parallel_size=2,
53
+ enforce_eager=True)
54
outputs = llm.generate(prompts, sampling_params)
55
# Print the outputs.
56
for output in outputs:
57
prompt = output.prompt
58
generated_text = output.outputs[0].text
57
- print(
58
- f"DP rank {dp_rank}, Prompt: {prompt!r}, "
59
- f"Generated text: {generated_text!r}")
59
+ print(f"DP rank {dp_rank}, Prompt: {prompt!r}, "
60
+ f"Generated text: {generated_text!r}")
61
62
63
if __name__ == "__main__":
vllm/config.py
+1
-1
@@ -1372,7 +1372,7 @@ class ParallelConfig:
1372
1373
@staticmethod
1374
def has_unfinished_dp(dp_group: "ProcessGroup",
1375
- has_unfinished: bool) -> bool:
1375
+ has_unfinished: bool) -> bool:
1376
tensor = torch.tensor([has_unfinished],
1377
dtype=torch.int32,
1378
device="cpu")
vllm/utils.py
+1
@@ -518,6 +518,7 @@ def get_open_port() -> int:
518
return port
519
return _get_open_port()
520
521
+
522
def _get_open_port() -> int:
523
port = envs.VLLM_PORT
524
if port is not None:
vllm/v1/engine/core_client.py
+2
-1
@@ -89,7 +89,7 @@ class EngineCoreClient(ABC):
89
90
def execute_dummy_batch(self) -> None:
91
raise NotImplementedError
92
-
92
+
93
async def execute_dummy_batch_async(self) -> None:
94
raise NotImplementedError
95
@@ -343,6 +343,7 @@ class SyncMPClient(MPClient):
343
def execute_dummy_batch(self) -> None:
344
self._call_utility("execute_dummy_batch")
345
346
+
347
class AsyncMPClient(MPClient):
348
"""Asyncio-compatible client for multi-proc EngineCore."""
349
vllm/v1/worker/gpu_model_runner.py
+2
-1
@@ -1167,7 +1167,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
1167
for k, v in self.intermediate_tensors.items()
1168
})
1169
1170
- with set_forward_context(None, self.vllm_config, num_tokens=num_tokens):
1170
+ with set_forward_context(None, self.vllm_config,
1171
+ num_tokens=num_tokens):
1172
hidden_states = model(
1173
input_ids=input_ids,
1174
positions=positions,