You can deploy this model and start earning money today!

Mahesh111000/
hanabi-qwen3-32b-lora-step65

Name: hanabi-qwen3-32b-lora-step65
Availability: InStock

Views

No views yet

Hanabi Qwen3-32B LoRA — step 65 (RL / GRPO via prime-rl)

RL training state for resuming on another machine. Base model: Qwen/Qwen3-32B (bf16).

Note: 97.5% of checkpoint_step_65/ is the frozen base model stored as fp32 master weights (32.76B x 4 bytes = 131 GB). Only ~3.2 GB is unique (LoRA + Adam moments). If you don't need the exact optimizer state and step counter, just use adapter/ (1.07 GB) with a fresh optimizer — it holds all the learning, and it lets you change the learning rate (a checkpoint resume does not).

path	size	purpose
`adapter/`	1.1 GB	LoRA adapter (r=32, alpha=64, fp32). The entire trained delta. PEFT/vLLM loadable.
`checkpoint_step_65/`	126 GB	prime-rl DCP training checkpoint — model + optimizer + scheduler, for an exact resume.
`code/`	small	the custom reward env, config, and launcher (not on PyPI — required).

Training state at step 65

67 steps completed (0–66); checkpoint saved at step 65.
LoRA r=32, alpha=64, lr 1e-5, all linear projections.
batch 256 = 16 problems x 16 rollouts, max_tokens 26624, seq_len 32768.
GPUs: 4 inference (TP=4) + 4 trainer (cp=4), 8x80GB.
Reward = deduction_score + move_reward (range [0,2]).

Resume on a new server

bash

git clone https://github.com/PrimeIntellect-ai/prime-rl.git
cd prime-rl && git checkout 16e747c25e81c5c90bd860c68a588f7b151353a6
uv sync

huggingface-cli download Mahesh111000/hanabi-qwen3-32b-lora-step65 --local-dir ./hanabi-state

# custom reward env (required — not on PyPI)
uv pip install --python .venv/bin/python --no-deps -e ./hanabi-state/code/hanabi-deduction

cp ./hanabi-state/code/rl_32b.toml configs/hanabi_deduction/
mkdir -p outputs/checkpoints
cp -r ./hanabi-state/checkpoint_step_65 outputs/checkpoints/step_65

# FLASHINFER = free 1.26x (measured, distribution-exact). vLLM defaults to the slower FlashAttention.
VLLM_ATTENTION_BACKEND=FLASHINFER   uv run rl @ configs/hanabi_deduction/rl_32b.toml --ckpt.resume-step 65

Dataset: Mahesh111000/Hanabi-init-30turns (or point dataset_path in the config at a local copy).

Expect: the first step after any (re)start is a cold start (async pipeline refills from empty) — 2.5–4 h. Steady state ~85–130 min/step. Needs 8x80 GB GPUs.

Full code, configs and profiling tooling: https://github.com/kaousheik-26/hanabi_latest_train_32b

Gotchas baked into the config (learned the hard way)

dist_timeout_seconds = 18000 — the default 600s kills the trainer on the cold-start step.
[orchestrator.client] timeout = 7200 — the default 1200s silently drops long rollouts.
ckpt.keep_last = 2 — without it, checkpoints fill the disk (~130 GB per save).
[trainer.tokenizer] name pinned — prime-rl otherwise resolves it from a stale default.
Resume always inherits the checkpoint's LR; the config's lr is ignored on resume. To change the learning rate you must start fresh from merged weights, not resume.
prime-rl's outputs/weights/ are NOT merged (bitwise identical to base) — don't rely on them.

Mahesh111000/hanabi-qwen3-32b-lora-step65

Hanabi Qwen3-32B LoRA — step 65 (RL / GRPO via prime-rl)

Contents

Training state at step 65

Resume on a new server

Gotchas baked into the config (learned the hard way)

Mahesh111000/
hanabi-qwen3-32b-lora-step65