reports/fp8-rl-grpo.md

RL on a 5090: FP8 RL hit a triple wall, so portable GRPO shipped a real +7.66

Goal: the marquee fine-tuning frontier — FP8 reinforcement learning (GRPO) on Blackwell, the one path most likely to hit the same toolchain wall that blocked vLLM inference (t036). Model: unsloth/Qwen3-4B + QLoRA GRPO adapter (r=32) · Reward: correctness (answer match) + format (#### <answer> marker) Hardware: RTX 5090 32GB · ~/grpo-env (Unsloth 2026.6.1, torch 2.10+cu128, sm_120) · Donald drained for the run, restored after Eval: in-process GSM8K, N=300, greedy, identical prompt + extraction for base and tuned

Result

	GSM8K (n=300)
Base (Qwen3-4B)	60.67% (182/300)
+ QLoRA GRPO (correctness + format reward)	68.33% (205/300)
Gain	+7.66 pts (+23 questions)

GRPO mean reward rose 1.375 → 1.725 over 150 steps (num_generations=8) — still climbing, not saturated. A clean, measured RL lift on a 4B, bigger than last week's +4.67 SFT gain on a 1B. But the headline isn't the number — it's which door was locked.

The marquee path was FP8 RL. It hit a triple wall.

The plan was FP8 RL: Unsloth exposes it as FastLanguageModel.from_pretrained(..., load_in_fp8=True), routing rollouts through vLLM for ~1.4x faster RL inference. On Blackwell, that path is blocked three independent ways — discovered in order:

1. The t036 signature, again — flashinfer FP8 kernels need CUDA ≥ 12.9. load_in_fp8=True loads unsloth/Qwen3-4B-FP8 and inits vLLM with quantization=fp8. vLLM logs SM 12.x requires CUDA >= 12.9; capsule's toolkit is 12.8 (nvcc 12.8.61), so flashinfer can't build the FP8 GEMM kernels. The exact wall t036's quant inference hit — now extended to RL.

2. An unsloth ↔ vLLM LoRA-API version crash. FP8 RL needs vLLM fast_inference + enable_lora=True. The only Blackwell-wheel vLLM (0.21.0) forces an older unsloth (2026.3.11), and that pair dies at LoRA warmup: 'LRUCacheWorkerLoRAManager' object has no attribute 'get_dummy_lora_warmup_rank' — an internal version split (unsloth 2026.3.11 + unsloth_zoo 2026.6.1).

3. An irreconcilable torch pin. Proven unsloth (2026.6.1) requires torch 2.10; vLLM 0.21 requires torch 2.11+cu130. They cannot share one env. And the unsloth version vLLM does allow (2026.3.11) has a broken GRPO loss path (NameError: align_completion_tool_mask).

Verdict: FP8 RL on this box is blocked by both a toolkit limit (CUDA 12.8 < 12.9) and a torch/unsloth/vLLM version knot. Same call as t036 — deferred, no toolkit install for one comparison.

What shipped instead: portable GRPO (use_vllm=False)

Rather than chase a CUDA toolkit, GRPO ran on plain HF generation — vLLM-free, so it sidesteps all three walls at once. Qwen3-4B 4-bit QLoRA policy, two rewards (correctness via the same GSM8K extractor the eval uses, plus a ####-format reward), the same prompt as the eval (train↔eval alignment — last week's hardest-won lever). ~18.8s/step at num_gen=4; the full run used num_gen=8, 150 steps. Slower than a vLLM rollout would be — but it produces the gain on hardware where the fast path won't load.

Worth it if / not if

• The portable path is the pragmatic one. If your box lacks a CUDA toolkit (or you just don't want the rabbit hole), use_vllm=False GRPO trades rollout speed for a stack that actually runs. The +7.66 is real either way.
• FP8 RL needs CUDA ≥ 12.9 on Blackwell. Not the model's fault, not Unsloth's — it's flashinfer's kernel build. Plan for a toolkit install if FP8 rollouts are the point.
• Not a capability miracle. GRPO on a 4B with a correctness+format reward sharpens what the model can already nearly do (reward climbed, didn't explode). A few points of measured GSM8K, not a tier jump.

Honest scope

In-process eval (internally consistent: base vs tuned, identical harness, raw generations saved for offline re-scoring) — not the leaderboard's server harness, so the absolute % isn't directly comparable to the model board. The delta is the claim.

Reproduce

# portable GRPO (no vLLM, no toolkit) — Qwen3-4B 4-bit QLoRA
~/grpo-env/bin/python scripts/train/gsm8k_grpo.py            # -> ~/unsloth-runs/gsm8k-grpo/adapter
~/grpo-env/bin/python scripts/train/eval_gsm8k.py --base unsloth/Qwen3-4B --out eval_base.json
~/grpo-env/bin/python scripts/train/eval_gsm8k.py --base unsloth/Qwen3-4B \
    --adapter ~/unsloth-runs/gsm8k-grpo/adapter --out eval_tuned.json

Scripts: scripts/train/{gsm8k_grpo,grpo_rewards,eval_gsm8k,gsm8k_extract}.py. Adapter: witcheer/qwen3-4b-gsm8k-grpo.