cc82f38b75
Merge 9cb07dd1ec into a3c7b01cc0
2026-09-10 21:11:19 -04:00
a3c7b01cc0
[fix] keep agent rollout token-aligned #850
2026-09-10 11:44:05 +08:00
9cb07dd1ec
perf: enable fused AdamW in pretrain training
...
Same fused AdamW optimization as the full SFT script, using the
existing device_type variable so the CPU-only path is unaffected.
Co-authored-by: TRAE CLI <traecli@bytedance.com>
2026-09-05 19:16:51 +08:00
859c8ee2a3
perf: enable fused AdamW in full SFT training
...
Use PyTorch's fused AdamW implementation to reduce optimizer kernel
launch overhead and speed up the training step on CUDA devices.
Guard with device check so the CPU-only fallback path is unaffected.
Co-authored-by: TRAE CLI <traecli@bytedance.com>
2026-09-05 18:37:04 +08:00
6fc918beb6
[update] remove expert args
2026-09-04 22:16:38 +08:00
4bd6062ead
Merge pull request #836 from JustinGuese/moe-cli-flags
...
Expose num_experts / num_experts_per_tok / seed on the trainer CLIs
2026-09-04 22:10:18 +08:00
4d7a74acab
Merge pull request #840 from JustinGuese/readme-benchmark-resolution
...
docs: state the measurement resolution of the objective benchmark table
2026-09-04 22:10:00 +08:00
77946fda00
[docs] 在客观评测表下补充各数据集的测量分辨率
...
lm_eval 对每个指标都会输出 stderr，但表格只保留了点估计。这里按测试集
大小给出各任务的 ± 标准误（obqa ±1.9 到 cmmlu ±0.4），并说明比较两个
模型时差值的标准误约为单项的 √2 倍，方便读者判断表中哪些差距是可以
解读的。中英文 README 同步。
2026-09-03 17:10:52 +07:00
fc585a9bf9
[update] 把随机种子暴露到命令行
...
trainer 里的种子此前硬编码为 42（setup_seed(42 + rank) 与每轮的
setup_seed(42 + epoch)），要跑多种子实验只能改源码。新增 --seed，
默认 42，与改动前完全一致。
2026-09-03 15:36:14 +07:00
c3199c7240
[update] 把 num_experts / num_experts_per_tok 暴露到命令行
...
此前要改专家数只能去编辑 model_minimind.py。三个入口
（train_pretrain / train_full_sft / eval_llm）的默认值与 MiniMindConfig
一致（4 / 1），默认行为不变；RL 与 agent 相关的 trainer 不动。
2026-09-03 15:13:29 +07:00
7a6fddd63a
[update] readme
2026-08-31 18:08:05 +08:00
c9a8ab4632
[update] dependency version
2026-08-31 17:54:37 +08:00
e84156818a
[update] dependency version
2026-08-29 15:12:28 +08:00
d65ef2c00e
[update] readme content
2026-08-21 16:49:00 +08:00
ed5ea9e02a
[update] readme content
2026-08-21 16:30:30 +08:00
eabad706da
[update] readme content
2026-08-21 16:25:20 +08:00
393e387e9a
[update] supporter charts
2026-08-06 22:07:29 +08:00
89d674b8a5
[refactor] PPO cleanup
2026-07-23 11:11:05 +08:00
2d4c385b0f
[fix] PPO rollout mode
...
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-07-23 11:05:51 +08:00
e9f46506da
Merge pull request #814 from FloatFrank/fix/ppo-logp-autocast-align-v3
...
[fix] PPO: align mb_resp_logp with old_resp_logp under autocast to stabilize training
2026-07-23 10:32:06 +08:00
1191fc8e9f
[fix] PPO: align mb_resp_logp with old_resp_logp under autocast
...
Move log_softmax for mb_resp_logp inside autocast_ctx so it is upcast
to fp32, matching the rollout path (compute_per_token_logps runs under
no_grad+autocast). Previously log_softmax ran outside autocast on
fp16/bf16 logits, introducing a ~1e-2 systematic bias between
mb_resp_logp and old_resp_logp, so ratio != 1 and KL != 0 even before
any weight update. Under sparse rule-based rewards (e.g. math PPO)
this bias destabilises training.
Also add:
- --rollout_eval_mode (default 1): switch actor to eval during rollout
to disable dropout and keep old/new logp consistent.
- --debug_log_ratio: optional diagnostic printing |mb-old| on the first
minibatch to verify ratio ~= 1.
2026-07-21 18:51:42 +08:00
512eed0b65
[fix] robustness
2026-06-28 18:03:34 +08:00
3f1a7cc25b
[fix] ddp exit hang
2026-06-01 17:50:39 +08:00
4a68da72d5
[fix] lora dims
2026-05-31 13:38:51 +08:00
4497610ec0
[fix] issue#771
2026-05-19 17:40:03 +08:00
9da8e1ab18
[update] requirements
2026-05-16 18:08:37 +08:00
dddedc6881
[fix] repetition_penalty
2026-05-07 19:08:52 +08:00
802c15b2b4
[feat] reduce RL memory
2026-05-06 15:07:28 +08:00
e73a407f7a
Merge pull request #759 from TKiteRunner/fix/grpo-oom-reorder-rewards
...
fix: resolve CUDA OOM in train_grpo.py on GPUs with <=8GB VRAM
2026-05-06 15:00:17 +08:00
10776417aa
fix: resolve CUDA OOM in train_grpo.py on GPUs with <=8GB VRAM
2026-05-06 14:14:15 +08:00
5020dc9dd4
[update] readme
2026-05-06 13:41:46 +08:00
bdee223036
[fix] inference bug
2026-05-03 20:48:48 +08:00
06d882e4ef
[update] readme
2026-05-01 12:06:47 +08:00
da865af63d
[update] readme
2026-04-28 17:22:25 +08:00
773e451b11
[fix] bugs
2026-04-27 19:16:08 +08:00
6361510016
[fix] rollout bugs
2026-04-27 17:54:09 +08:00
d4c6bc5c7e
[update] readme
2026-04-27 10:59:41 +08:00
24896cd2c4
[update] readme
2026-04-24 20:16:06 +08:00
e2fc397176
[update] readme
2026-04-24 20:14:10 +08:00
693fb1ccf1
[update] readme
2026-04-21 14:34:46 +08:00
5416a44471
[fix] bugs
2026-04-21 13:03:34 +08:00
1718e9a44d
[fix] transformers-5.x
2026-04-19 23:48:54 +08:00
5704766352
[update] tie embedding
2026-04-19 21:57:28 +08:00
1ea113ea2c
[update] readme
2026-04-19 14:51:47 +08:00
487f78754d
[update] readme
2026-04-10 10:55:24 +08:00
e796b8028a
[fix] lora compile
2026-04-10 00:01:21 +08:00
b2488e6440
[update] readme
2026-04-09 19:00:41 +08:00
939dc8ff42
[update] readme
2026-04-09 18:54:17 +08:00
cadacabecb
[update] bench
2026-04-09 17:12:48 +08:00
5351424bf0
[fix] lora moe
2026-04-09 16:36:48 +08:00
Neil Ge
jingyaogong
neil.ge
Justin Guese
FloatFrank
TKiteRunner
jingyaogong