node B vLLM 0.33.0 env for Llama-3.2-3B corpus generation (board 11-16 07:44): README.md
Browse files- env/vllm_0.33.0/README.md +22 -0
env/vllm_0.33.0/README.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# vLLM 0.33.0 environment used for the L1 corpus regeneration on node B (11-15/16 UTC)
|
| 2 |
+
|
| 3 |
+
For node D2's Llama-3.2-3B corpus generation (board user decision 11-16 07:44 UTC, item (3)).
|
| 4 |
+
|
| 5 |
+
- `vllm_env_0.33.0_py312_cu130.tar.zst`: the venv as built on node B (`uv venv --python 3.12` + `uv pip install vllm`).
|
| 6 |
+
Extract at the same path: `mkdir -p /workspace && tar -C /workspace -I zstd -xf vllm_env_0.33.0_py312_cu130.tar.zst`
|
| 7 |
+
(creates `/workspace/vllm_env`). Its interpreter links to `/usr/local/bin/python3.12` (Python 3.12.3); the node needs
|
| 8 |
+
that Python. Versions: vllm 0.33.0, torch 2.14.0 (cu130), transformers 5.20.2, flashinfer-python 0.7.2.post1.
|
| 9 |
+
Driver on node B: 595.91.07 (CUDA 13.2).
|
| 10 |
+
- `requirements_frozen.txt`: `uv pip freeze` of the same venv, to rebuild instead of extracting
|
| 11 |
+
(`uv venv --python 3.12 /workspace/vllm_env && VIRTUAL_ENV=/workspace/vllm_env uv pip install -r requirements_frozen.txt`).
|
| 12 |
+
|
| 13 |
+
**Required at run time: `VLLM_USE_FLASHINFER_SAMPLER=0`.** Without it every shard fails at engine start-up: vLLM's
|
| 14 |
+
FlashInfer sampler JIT-builds a kernel and needs `ninja` and `nvcc`, which node B does not have (node B's first L1 data run
|
| 15 |
+
on 11-15 19:23 failed this way). Greedy (T = 0) generation needs no FlashInfer sampler.
|
| 16 |
+
|
| 17 |
+
Generator: `gen_t0.py` from `origin/icml27/sat-20261106` (`icml27/serial/gen_t0.py`, unchanged), run per shard as
|
| 18 |
+
`CUDA_VISIBLE_DEVICES=<g> HF_HUB_OFFLINE=1 VLLM_USE_FLASHINFER_SAMPLER=0 VLLM_NO_USAGE_STATS=1 DO_NOT_TRACK=1 /workspace/vllm_env/bin/python -I gen_t0.py --target <model dir> --split <split json> --shard <i>/<n> --out <file>`
|
| 19 |
+
(node B's chain: `div2/icml27/...`; local copy `/workspace/opt_runs/icml27_b_l1data/chain_l1data.sh`). On node B, 24,249
|
| 20 |
+
Qwen3-8B prompts (4 shards on 4 x H200) took about 4 minutes.
|
| 21 |
+
|
| 22 |
+
No tokens or credentials are inside the tarball (checked for the HF and GitHub token strings before packing).
|