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, 17:15 UTC
[TPU] [V1] fix cases when max_num_reqs is set smaller than MIN_NUM_SEQS (#15583)
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Chengji Yao committed
Mar 26, 2025 at 22:46 UTC
619d3de8bd54be017d8d8211259ea6ad4865ecbe
2 files changed
+2
-5
examples/offline_inference/tpu.py
+1
-4
@@ -14,10 +14,7 @@ answers = [
14
]
15
N = 1
16
# Currently, top-p sampling is disabled. `top_p` should be 1.0.
17
-sampling_params = SamplingParams(temperature=0.7,
18
- top_p=1.0,
19
- n=N,
20
- max_tokens=16)
17
+sampling_params = SamplingParams(temperature=0, top_p=1.0, n=N, max_tokens=16)
18
19
# Set `enforce_eager=True` to avoid ahead-of-time compilation.
20
# In real workloads, `enforace_eager` should be `False`.
vllm/v1/worker/tpu_model_runner.py
+1
-1
@@ -88,7 +88,7 @@ class TPUModelRunner:
88
self.max_model_len = model_config.max_model_len
89
self.max_num_blocks_per_req = cdiv(self.max_model_len, self.block_size)
90
self.max_num_tokens = scheduler_config.max_num_batched_tokens
91
- self.max_num_reqs = scheduler_config.max_num_seqs
91
+ self.max_num_reqs = max(scheduler_config.max_num_seqs, MIN_NUM_SEQS)
92
93
# Model-related.
94
self.num_attn_layers = model_config.get_num_layers_by_block_type(