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, 09:30 UTC
[TPU] Correctly profile peak memory usage & Upgrade PyTorch XLA (#9438)
Woosuk Kwon committed
Oct 30, 2024 at 02:41 UTC
211fe91aa88730c04df439298d8103a587302493
3 files changed
+11
-10
Dockerfile.tpu
+1
-1
@@ -1,4 +1,4 @@
1
-ARG NIGHTLY_DATE="20240828"
1
+ARG NIGHTLY_DATE="20241017"
2
ARG BASE_IMAGE="us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm_$NIGHTLY_DATE"
3
4
FROM $BASE_IMAGE
docs/source/getting_started/tpu-installation.rst
+2
-2
@@ -56,8 +56,8 @@ First, install the dependencies:
56
$ pip uninstall torch torch-xla -y
57
58
$ # Install PyTorch and PyTorch XLA.
59
- $ export DATE="20240828"
60
- $ export TORCH_VERSION="2.5.0"
59
+ $ export DATE="20241017"
60
+ $ export TORCH_VERSION="2.6.0"
61
$ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-${TORCH_VERSION}.dev${DATE}-cp310-cp310-linux_x86_64.whl
62
$ pip install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-${TORCH_VERSION}.dev${DATE}-cp310-cp310-linux_x86_64.whl
63
vllm/worker/tpu_worker.py
+8
-7
@@ -133,18 +133,19 @@ class TPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
133
# Synchronize before measuring the memory usage.
134
xm.wait_device_ops()
135
136
- dtype_btyes = get_dtype_size(self.cache_dtype)
137
- block_size = self.cache_config.block_size
138
- block_size_bytes = (dtype_btyes * block_size * num_layers * 2 *
139
- head_size * num_kv_heads)
140
-
141
- # Calculate the TPU KV cache size based on profiling.
136
+ # Get the maximum amount of memory used by the model weights and
137
+ # intermediate activations.
138
m = xm.get_memory_info(self.device)
139
total_memory_size = m["bytes_limit"]
140
+ profiled = m["peak_bytes_used"] # Weights + intermediate activations.
141
+
142
+ # Calculate the TPU KV cache size based on profiling.
143
usable_memory_size = int(total_memory_size *
144
self.cache_config.gpu_memory_utilization)
146
- profiled = m["bytes_used"] # Weights + intermediate activations.
145
tpu_kv_cache_bytes = max(usable_memory_size - profiled, 0)
146
+ dtype_btyes = get_dtype_size(self.cache_dtype)
147
+ block_size_bytes = (dtype_btyes * self.cache_config.block_size *
148
+ num_layers * 2 * head_size * num_kv_heads)
149
num_tpu_blocks = tpu_kv_cache_bytes // block_size_bytes
150
num_tpu_blocks = (num_tpu_blocks // 8) * 8 # Round down to 8.
151