New scripts under scripts/benchmarks/:
- flash_attn_fa4_shim.py: monkey-patches that let transformers CB dispatch to
Flash Attention 4 on Blackwell (sm_100). CB's ContinuousBatchProcessor
otherwise emits a 4D paged attention mask for flash_attention_2 (which then
breaks _flash_attention_forward's _upad_input branch), and passes
max_seqlen_q instead of max_length_q. The shim skips the mask for FA and
accepts both names.
- persistent_cb.py: replaces model.generate_batch with a version that reuses
one ContinuousBatchingManager across calls, avoiding the per-step
PagedAttentionCache realloc. Wired up behind --persistent_cb on the tpaged
and standalone scripts.
- qwen3_grpo_naive.py: vanilla HF model.generate + TRL GRPOTrainer, no vLLM
and no CB. Mirrors the TRL docs example. Useful as a third column in the
comparison and also as a "will this at least converge" sanity check.
Adds --attn_impl and --persistent_cb flags to the existing generation and
training scripts. No changes to Unsloth internals.
Updated README.md with the FA install recipe (flash-attn-4==4.0.0b9, plus
a small site-packages shim that re-exports FA4's cute.* symbols under the
FA2 flash_attn namespace so transformers' is_flash_attn_2_available() and
_lazy_imports("flash_attention_2") succeed on B200).
Benchmark numbers on a single B200, Qwen3-4B-Base LoRA rank 32, bf16:
Generation microbenchmark (32 prompts, 512 new tokens):
vLLM 7224 decode tok/s (100%)
CB paged|sdpa 527 decode tok/s ( 7.3%)
CB paged|flash_attention_2 (FA4) 709 decode tok/s ( 9.8%)
CB paged|flash_attention_2 persistent 529 decode tok/s ( 7.3%)
GRPO training (max_steps=20, num_generations=2, per_device_batch=2):
vLLM colocated 136.6 s peak 157 GB
naive TRL (HF generate) 910.0 s peak 15 GB
CB SDPA 1521.5 s peak 98 GB (prior run)
CB FA4 1470.7 s peak 82 GB
CB FA4 + persistent 1562.1 s peak 87 GB
CB FA4 + ng=4 persistent 1597.0 s peak 94 GB
FA4 is a real ~1.4x improvement over SDPA for CB decode throughput but the
50% of vLLM target is still not reached. The remaining gap is driven by
CUDA graph capture (which ContinuousBatchingManager still NotImplementedErrors
on) and vLLM's scheduler being more efficient for decode-heavy GRPO rollouts.
Naive TRL generate is the honest small-rig baseline: 6.7x slower than vLLM
at 10% of the VRAM footprint, and ~1.7x faster than CB here.
166 lines
8.2 KiB
Markdown
166 lines
8.2 KiB
Markdown
# Qwen3-4B GRPO rollout engine benchmarks
|
|
|
|
This directory holds the reproducible scripts behind the experiment
|
|
documented in the accompanying PR: can Hugging Face transformers'
|
|
continuous-batching API (`model.generate_batch`, backed by
|
|
`PagedAttentionCache`) serve as a drop-in replacement for vLLM during GRPO
|
|
rollouts on the Qwen3-4B notebook?
|
|
|
|
The short answer on a single NVIDIA B200 with Unsloth Qwen3-4B-Base, LoRA
|
|
rank 32, bf16: transformers continuous batching is functionally correct and
|
|
integrates with TRL's `use_transformers_paged=True` path, but end-to-end
|
|
throughput lands at around 7 to 10 percent of vLLM colocated even after
|
|
wiring in Flash Attention 4 on Blackwell. Full numbers and per-step timings
|
|
are in the PR description.
|
|
|
|
## Files
|
|
|
|
| File | Purpose |
|
|
|---|---|
|
|
| `unsloth_grpo_common.py` | Shared dataset loading, reward functions, and GRPO hyperparameters |
|
|
| `qwen3_grpo_vllm.py` | vLLM baseline training entry (`fast_inference=True`, `use_vllm=True`, `vllm_mode="colocate"`) |
|
|
| `qwen3_grpo_naive.py` | Naive TRL path (vanilla HF `model.generate`, no vLLM, no CB) matching https://huggingface.co/docs/trl/grpo_trainer |
|
|
| `qwen3_grpo_tpaged.py` | Continuous-batching candidate (`fast_inference=False`, `use_transformers_paged=True`, vanilla HF + PEFT). Supports `--persistent_cb` |
|
|
| `cb_vs_vllm_generation.py` | Standalone generation microbenchmark across both engines; supports `--attn_impl`, `--persistent_cb` |
|
|
| `flash_attn_fa4_shim.py` | Installs two monkey-patches that let CB dispatch to FA4 when `--attn_impl flash_attention_2` is selected |
|
|
| `persistent_cb.py` | Replaces `model.generate_batch` with a version that reuses a single `ContinuousBatchingManager` |
|
|
|
|
## Flash Attention 4 on Blackwell (sm_100)
|
|
|
|
The CB code path has three attention implementations: `eager_paged`,
|
|
`sdpa_paged`, `flash_attention_2`. The last one requires the legacy
|
|
`flash_attn` Python package, which does not install cleanly on B200 today:
|
|
|
|
- `flash_attn==2.8.3+cu12torch2.8cxx11abiTRUE-cp313` from the Dao-AILab
|
|
releases hits `undefined symbol: _ZNK3c106SymInt6sym_neERKS0_` on torch
|
|
2.9.1 (ABI drift between torch 2.8 and 2.9).
|
|
- `flash_attn_3-3.0.0-cp39-abi3-manylinux_2_28_x86_64.whl` from the PyTorch
|
|
wheel index installs but was built for sm_80 and sm_90a only. B200 is
|
|
sm_100. The kernel call fails with "no kernel image is available for
|
|
execution on the device".
|
|
- `flash-attn-4==4.0.0b9` (pure Python CuTeDSL, Dao-AILab) works on B200.
|
|
It exposes `flash_attn.cute.flash_attn_varlen_func`.
|
|
|
|
The recipe this repo uses:
|
|
|
|
```bash
|
|
uv pip install --no-deps flash-attn-4==4.0.0b9
|
|
```
|
|
|
|
plus a tiny site-packages shim that re-exports FA4 symbols under the
|
|
FA2 `flash_attn` namespace so transformers' `is_flash_attn_2_available()` and
|
|
`_lazy_imports("flash_attention_2")` succeed. The shim lives out of tree in
|
|
`lib/python3.13/site-packages/flash_attn/__init__.py` +
|
|
`flash_attn/bert_padding.py` + a `flash_attn-2.8.3.dist-info/` directory
|
|
with enough metadata to satisfy `importlib.metadata.version("flash_attn")`.
|
|
|
|
On top of that, `flash_attn_fa4_shim.py` monkey-patches two rough edges in
|
|
the CB to FA integration that are unrelated to which FA version you use:
|
|
|
|
1. `ContinuousBatchProcessor.return_attention_mask` returns `False` for
|
|
`flash_attention_2` / `flash_attention_3` so CB does not emit a 4-D paged
|
|
attention mask that breaks `_flash_attention_forward`'s `_upad_input`
|
|
branch.
|
|
2. `_flash_attention_forward` accepts `max_seqlen_q` / `max_seqlen_k`
|
|
as aliases for `max_length_q` / `max_length_k`. Without this rename CB's
|
|
model kwargs never bind and FA is called with `max_seqlen_q=None`.
|
|
|
|
## Reproduce
|
|
|
|
```bash
|
|
pip install unsloth "transformers>=4.57" "trl>=0.25" peft vllm
|
|
uv pip install --no-deps flash-attn-4==4.0.0b9
|
|
|
|
# Generation microbenchmark (32 prompts, 512 new tokens each)
|
|
CUDA_VISIBLE_DEVICES=2 python scripts/benchmarks/cb_vs_vllm_generation.py \
|
|
--backend vllm --stats_path logs/vllm_gen.json \
|
|
--n_prompts 32 --n_rounds 2 --max_new_tokens 512 \
|
|
--gpu_memory_utilization 0.6
|
|
|
|
# CB variants
|
|
CUDA_VISIBLE_DEVICES=6 python scripts/benchmarks/cb_vs_vllm_generation.py \
|
|
--backend tpaged --attn_impl sdpa \
|
|
--stats_path logs/cb_gen_sdpa.json \
|
|
--n_prompts 32 --n_rounds 2 --max_new_tokens 512
|
|
|
|
CUDA_VISIBLE_DEVICES=6 python scripts/benchmarks/cb_vs_vllm_generation.py \
|
|
--backend tpaged --attn_impl flash_attention_2 \
|
|
--stats_path logs/cb_gen_fa.json \
|
|
--n_prompts 32 --n_rounds 2 --max_new_tokens 512
|
|
|
|
# Full GRPO training (20 steps)
|
|
CUDA_VISIBLE_DEVICES=2 python scripts/benchmarks/qwen3_grpo_vllm.py \
|
|
--max_steps 20 --num_generations 2 --per_device_train_batch_size 2 \
|
|
--output_dir outputs/grpo_vllm --stats_path logs/vllm_stats.json \
|
|
--gpu_memory_utilization 0.6
|
|
|
|
CUDA_VISIBLE_DEVICES=7 python scripts/benchmarks/qwen3_grpo_naive.py \
|
|
--max_steps 20 --num_generations 2 --per_device_train_batch_size 2 \
|
|
--output_dir outputs/grpo_naive --stats_path logs/naive_stats.json
|
|
|
|
CUDA_VISIBLE_DEVICES=6 python scripts/benchmarks/qwen3_grpo_tpaged.py \
|
|
--max_steps 20 --num_generations 2 --per_device_train_batch_size 2 \
|
|
--attn_impl flash_attention_2 \
|
|
--output_dir outputs/grpo_tpaged_fa --stats_path logs/tpaged_stats_fa.json \
|
|
--max_batch_tokens 16384 --num_blocks 16384
|
|
```
|
|
|
|
## Known integration notes for transformers continuous batching + TRL + Unsloth
|
|
|
|
These are the sharp edges you hit going down the continuous-batching path and
|
|
how `qwen3_grpo_tpaged.py` handles them:
|
|
|
|
1. **`top_k=-1` is not a valid value for transformers.** vLLM treats `-1` as
|
|
"disabled", but `TopKLogitsWarper` raises
|
|
`ValueError: top_k has to be a strictly positive integer`. The script
|
|
rewrites `top_k=None` on the shared GRPOConfig before handing it to
|
|
`GRPOConfig(use_transformers_paged=True, ...)`.
|
|
|
|
2. **`PagedAttentionCache` default upper bounds are extremely conservative.**
|
|
`_upper_bound_max_batch_tokens=256` and `_upper_bound_num_blocks=4096`
|
|
choke decode throughput. The script passes
|
|
`generation_kwargs={"max_batch_tokens": 16384, "num_blocks": 16384}` which
|
|
TRL forwards to `GenerationConfig`, and the CB manager reads them when
|
|
sizing the paged cache.
|
|
|
|
3. **Unsloth's `Qwen3Attention_fast_forward` bypasses the functional
|
|
attention interface.** Calling `model.generate_batch` on an
|
|
Unsloth-patched Qwen3 model fails inside
|
|
`unsloth.utils.attention_dispatch.run_attention` because Unsloth routes
|
|
through its own dispatcher rather than reading
|
|
`config._attn_implementation`. The benchmark script works around this by
|
|
loading a vanilla HF Qwen3 with PEFT LoRA for the tpaged and naive paths.
|
|
This costs the Unsloth training kernels but keeps the comparison clean.
|
|
A proper upstream fix is to detect `config._attn_implementation` being
|
|
`flash_attention_2` / `sdpa_paged` / `eager_paged` and delegate to the
|
|
stock transformers forward.
|
|
|
|
4. **TRL imports `GuidedDecodingParams` from `vllm.sampling_params`.**
|
|
Newer vLLM releases (>= 0.13) have moved or removed that symbol, so
|
|
`trl.trainer.grpo_trainer` fails to import on a fresh vLLM install even
|
|
if you are not using vLLM. `qwen3_grpo_tpaged.py` installs a minimal
|
|
shim before importing TRL.
|
|
|
|
5. **`UnslothGRPOTrainer` calls `model.for_training()` /
|
|
`for_inference()`.** Importing `unsloth` replaces
|
|
`trl.GRPOTrainer` with `UnslothGRPOTrainer`, which assumes the model has
|
|
these hooks. A vanilla HF model does not, so
|
|
`qwen3_grpo_tpaged.py` does not `import unsloth` at all.
|
|
|
|
## Why continuous batching is still slower than vLLM on this workload
|
|
|
|
- `ContinuousBatchingManager` does not yet implement CUDA graphs
|
|
(`use_cuda_graph=True` raises `NotImplementedError`). vLLM captures 100+
|
|
mixed prefill-decode and decode graphs during warmup.
|
|
- CB re-allocates a fresh `PagedAttentionCache` on every `generate_batch`
|
|
call. For GRPO that is once per step. `--persistent_cb` (via
|
|
`persistent_cb.py`) keeps the cache warm across steps.
|
|
- FA4 is a CuTeDSL package: first call per shape pays a one-time
|
|
JIT-compile.
|
|
- vLLM uses its own colocated attention + FlashInfer / TRTLLM kernels
|
|
tuned for decode, which currently outperform everything a generic CB
|
|
path can do.
|
|
|
|
These are all upstream transformers issues, not Unsloth issues. The scripts
|
|
in this directory are intentionally simple so they are easy to port into a
|
|
future upstream fix.
|