Qwen3-8B EAGLE3 example: learning_rate=1e-4 shows large grad-norm outliers in early training

Hi, thanks for releasing SpecForge and the Qwen3-8B EAGLE3 training example.

• I have searched related issues but did not find the same LR / grad_norm issue.
• I have checked the current example script and this still applies to the latest version I tested.
• I am providing the environment, reproduction setup, logs/metrics, and a small decode-style check.
• This is a bug / training-recipe stability report rather than a general usage question.
• I am using English for this issue.

I ran a learning-rate ablation based on the current examples/run_qwen3_8b_eagle3_online.sh recipe and observed that the default --learning-rate 1e-4 may be too aggressive for the Qwen3-8B EAGLE3 setting, at least during early training.

This is not a full convergence study. Due to compute and disk limits, I only trained for about one epoch plus a small part of the second epoch. The main signal I want to report is the logged grad_norm: with the official 1e-4 LR, the run shows many very large gradient-norm outliers in the middle of training, while smaller learning rates are much more stable under the same setup.

Setup

Base recipe:

• Script: examples/run_qwen3_8b_eagle3_online.sh
• Target model: Qwen3-8B
• Draft config: configs/qwen3-8b-eagle3.json
• Dataset: ShareGPT-style training data, following the official example's sharegpt_train.jsonl setup
• num_epochs=10
• batch_size=1
• max_length=4096
• chat_template=qwen
• target_model_backend=sglang
• warmup_ratio=0.015
• max_grad_norm=0.5
• ttt_length=7
• tp_size=1

The only intended ablation variable was LR:

• 1e-4
• 5e-5
• 2e-5
• 1e-5

One implementation detail: I used --attention-backend fa for the draft model because the default flex_attention path hit a Triton resource-limit compile error on our RTX A6000 machine. All LR runs used the same FA2 backend, so this should not explain the difference between LR settings.

Training length

The ShareGPT-style training file contained 68,623 rows, so one epoch is about 68.6k steps with batch_size=1.

The runs reached about global_step=75,999 before stopping because the disk filled while saving a checkpoint. The last complete checkpoint was global_step=75,000; the last checkpoint before the end of the first epoch was epoch_0_step_68000.

Main observation: logged grad_norm outliers

The 1e-4 run produced many large logged grad_norm outliers. Smaller learning rates were much more stable.

Stats up to global_step=75,000:

LR	median grad_norm	p95	p99	max	count grad_norm > 100	count grad_norm > 200
1e-4	10.79	53.74	144.71	5900.89	1316	434
5e-5	9.10	26.26	36.33	218.68	6	1
2e-5	14.76	33.37	43.92	443.87	13	2
1e-5	17.72	38.33	49.50	407.52	18	2

Visually, the 1e-4 run has a large cluster of grad-norm spikes around the middle of the first epoch, while the lower-LR runs remain much smoother.

Training metrics near the last complete checkpoint

At global_step=75,000, the lower-LR runs also had better training metrics:

LR	loss	acc	grad_norm
1e-4	1.85	0.51	11.22
5e-5	1.40	0.60	12.96
2e-5	1.43	0.60	19.66
1e-5	1.64	0.56	23.69

Among the LRs I tried, 5e-5 looked best overall. 2e-5 was close in training loss/acc but slightly less favorable, and 1e-5 appeared slower. I do not claim 5e-5 is globally optimal, only that it looked clearly better than 1e-4 in this run.

Small GSM8K decode-style check

I also ran a small EAGLE3 teacher-forced / greedy-rollout check on the first 16 GSM8K test examples, formatted with the Qwen chat template.

For this check I used the last checkpoint before the end of the first epoch, epoch_0_step_68000. I could only evaluate 1e-4 and 5e-5 because I had deleted the 2e-5 and 1e-5 checkpoints during disk cleanup.

LR	teacher-forced accept len	greedy-rollout accept len	teacher-forced mean acc	greedy-rollout mean acc
1e-4	0.9367	0.8919	0.3589	0.2064
5e-5	1.6861	1.5612	0.5545	0.3122

Greedy-rollout per-position accuracy:

LR	greedy-rollout per-position acc
1e-4	[0.5706, 0.3225, 0.1946, 0.1300, 0.0864, 0.0706, 0.0701]
5e-5	[0.7310, 0.4846, 0.3388, 0.2365, 0.1664, 0.1292, 0.0992]

This small decode-style check is consistent with the training curves: 5e-5 looked substantially better than 1e-4 at approximately the same training stage.

Question / suggestion

Is 1e-4 the intended LR for Qwen3-8B EAGLE3, or could it be too high for this example?

Would you consider lowering the default LR in examples/run_qwen3_8b_eagle3_online.sh, or documenting that users may want to try a smaller LR such as 5e-5 for Qwen3-8B?

Again, this is only an early-training ablation and not a full best-LR sweep. I am mainly reporting that 1e-4 shows unusually large grad-norm outliers under a setup close to the official example, while smaller LRs look more stable and 5e-5 also gives better early decode-style metrics.

Start with examples/run_qwen3_8b_eagle3_online.sh and configs/qwen3-8b-eagle3.json, then reproduce the reported learning-rate comparison with the stated training settings and inspect grad_norm and checkpoint metrics. Compare 1e-4 with the lower learning rates, including the decode-style check if checkpoints are available; done means establishing whether the example should change its default or document a smaller recommended LR.