Downloads · 30 days
0
jhhj25/reef-nccl-ep-qwen3-8b-lora-16k-run2
reef-nccl-ep-qwen3-8b-lora-16k-run2 is a reinforcement learning model from jhhj25. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
LoRA adapters (rank 32, base Qwen3-8B) from a 16-step run of grouped relative-reward RL on the 16384-token variant of an NCCL EP kernel-tuning task. All 16 intermediate checkpoints are included as step0 ... step15, ea…
Downloads · 30 days
0
Access
Public
Updated Sep 6, 2026
Repo size
2.8 GB
Likes
0
Public
Click a slice to open those files.
.safetensors2.8 GB · 100%
From the Hugging Face model README
LoRA adapters (rank 32, base Qwen3-8B) from a 16-step run of grouped
relative-reward RL on the 16384-token variant of an NCCL EP kernel-tuning
task. All 16 intermediate checkpoints are included as step_0 ... step_15,
each containing adapter_config.json and adapter_model.safetensors.
The 16384-token variant of the kernel-tuning task had never been tuned before these runs; only an untuned chunk=512 anchor configuration existed at this size. Two 16-step runs of grouped relative-reward RL (reef TTTD with slime/Megatron LoRA training) produced 28 measured configurations at this size, all of them model-original discoveries.
The best configuration found is qps=2 chunk=512 subputs=8 cb=32 sms=11,
which reaches 191.45 GB/s with correctness 16/16. Over training, the runs
converged to a sampling attractor in the 189.5-190.9 GB/s range.
The optimum previously found at 8192 tokens (subputs=6 cb=32 sms=11 on
chunk=256) transferred structurally to this size: at 16384 tokens the same
subputs/cb/sms region wins, applied to the chunk=512 anchor.
The stored adapter_config.json records a local placeholder as the base
model path, so pass the base model explicitly:
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
model = PeftModel.from_pretrained(base, "jhhj25/reef-nccl-ep-qwen3-8b-lora-16k-run2", subfolder="step_15")
step_15 is the final checkpoint; earlier steps trace the training curve.