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, 07:15 UTC
[TPU] update torch_xla pin (#19231)
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Chengji Yao committed
Jun 5, 2025 at 21:27 UTC
b61dc5f972b3130ab6a3bc0fb57b990d02a3f8fb
3 files changed
+8
-7
requirements/tpu.txt
+5
-5
@@ -18,9 +18,9 @@ setuptools==78.1.0
18
--find-links https://storage.googleapis.com/libtpu-releases/index.html
19
--find-links https://storage.googleapis.com/jax-releases/jax_nightly_releases.html
20
--find-links https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
21
-torch==2.8.0.dev20250529
22
-torchvision==0.22.0.dev20250529
23
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250529-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
24
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250529-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
25
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250529-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
21
+torch==2.8.0.dev20250605
22
+torchvision==0.23.0.dev20250605
23
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250605-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
24
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250605-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
25
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250605-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
26
tests/tpu/test_moe_pallas.py
+1
-1
@@ -27,7 +27,7 @@ TOP_KS = [2, 6]
27
# The Pallas GMM kernel requires num_tokens * topk to be a multiple of 16
28
@pytest.mark.parametrize("m", [8, 16, 64, 2048])
29
@pytest.mark.parametrize("n", [128, 1024, 2048])
30
-@pytest.mark.parametrize("k", [128, 512, 1024])
30
+@pytest.mark.parametrize("k", [128, 511, 1024])
31
@pytest.mark.parametrize("e", NUM_EXPERTS)
32
@pytest.mark.parametrize("topk", TOP_KS)
33
@pytest.mark.parametrize("ep_size", EP_SIZE)
vllm/v1/worker/tpu_worker.py
+2
-1
@@ -100,7 +100,8 @@ class TPUWorker:
100
# `xla_tpu_force_1d_allreduce_at_chunk_count` is a temporary solution to
101
# fix this. It will be removed after the bug in XLA compiler is fixed.
102
os.environ["LIBTPU_INIT_ARGS"] = (
103
- "--xla_tpu_force_1d_allreduce_at_chunk_count=1")
103
+ os.environ.get("LIBTPU_INIT_ARGS", "") +
104
+ " --xla_tpu_force_1d_allreduce_at_chunk_count=1")
105
torch.set_grad_enabled(False)
106
torch.set_default_dtype(self.model_config.dtype)
107