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 16, 03:15 UTC
Update requirements-tpu.txt to support python 3.9 and 3.11 (#11695)
Signed-off-by: mgoin <michael@neuralmagic.com>
Michael Goin committed
Jan 3, 2025 at 17:36 UTC
bf0d97d78619b290ed273199ad3800b57b638603
2 files changed
+7
-3
requirements-tpu.txt
+3
-1
@@ -18,6 +18,8 @@ ray[default]
18
--find-links https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
19
torch==2.6.0.dev20241126+cpu
20
torchvision==0.20.0.dev20241126+cpu
21
-torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241126-cp310-cp310-linux_x86_64.whl
21
+torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241126-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
22
+torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241126-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
23
+torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241126-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
24
jaxlib==0.4.36.dev20241122
25
jax==0.4.36.dev20241122
vllm/worker/tpu_model_runner.py
+4
-2
@@ -126,8 +126,10 @@ class TPUModelRunner(ModelRunnerBase[ModelInputForTPU]):
126
logger.warning(
127
"The max_model_len (%d) is too large. This may degrade the "
128
"performance due to the insufficient smem size. Consider "
129
- "setting --max-model-len to a smaller value.",
130
- self.model_config.max_model_len)
129
+ "setting --max-model-len to a smaller value, like %d.",
130
+ self.model_config.max_model_len,
131
+ self.model_config.max_model_len /
132
+ (block_table_size / smem_size))
133
134
def load_model(self) -> None:
135
self.device = self.device_config.device