c54a88887b
Merge 6b2f4c939a into 7a9137d2e9
2026-09-15 16:19:57 +08:00
0f6803e551
[fix] replace tool eval() with an AST math evaluator
...
execute_tool() relied on signal.SIGALRM, which does not exist on Windows.
The AttributeError was swallowed by a bare except, so every tool call
silently returned None and the model always observed
{"error": "tool not found"} (issue #846 ).
calculate_math also ran eval() on model-generated text. Two of the three
call sites (scripts/eval_toolcall.py, scripts/web_demo.py) passed no
globals at all, leaving full builtins reachable from sampled output.
safe_math_eval() in trainer_utils.py parses the expression with ast and
walks it, allowing only arithmetic operators and a math whitelist, with
caps on expression length and exponent size. It cannot loop forever, so
the signal-based timeout is no longer needed and is removed. Both sqrt(x)
and math.sqrt(x) spellings keep working.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-07 07:53:33 +08:00
6b2f4c939a
[update] MoE 分发：每层主机同步从 3E 次降到 1 次
...
专家循环里 mask.any() / nonzero() / 布尔索引每个专家各触发一次 device→host
同步，实测每层正好 3E 次（默认 4 专家 × 8 层 = 每次前向 96 次）。同步会清空
CUDA 流，而这些开销全部花在路由的簿记上。
改为先按专家把 token 排成连续段，用 searchsorted 取边界、一次 tolist() 读回
主机，每个专家取一段连续切片。专家循环保留，整层只剩 1 次同步，与专家数无关。
纯 PyTorch，无新依赖，对硬件和 dtype 没有要求。
stable=True 保证段内 token 顺序不变，因此前向与全部参数梯度在所有 k 下与原
写法逐位相同；输入梯度在 k<=2 时逐位相同，k>=3 时因浮点加法不满足结合律
差约 1 ulp（相对 1e-7）。原掩码写法作为参考实现保留在 scripts/test_moe_sorted.py，
每次跑测试都会执行并断言一致。
state_dict、aux_loss、router 与模块结构均未改动，旧 checkpoint 仍可 strict 加载。
2026-09-03 15:05:25 +07:00
512eed0b65
[fix] robustness
2026-06-28 18:03:34 +08:00
5704766352
[update] tie embedding
2026-04-19 21:57:28 +08:00
5351424bf0
[fix] lora moe
2026-04-09 16:36:48 +08:00
b7e0ae21d6
[update] default model
2026-03-31 13:40:16 +08:00
6b0b0c5e2f
[update] fp16 inference
2026-03-27 16:29:46 +08:00
101d7df2da
[update] minimind-3
2026-03-25 23:57:45 +08:00
04616c41a5
[update] safe half
2026-01-30 20:29:31 +08:00
07364c3fbe
[update] rename train tokenizer
2026-01-06 01:17:33 +08:00
a8455ca8a3
[fix] messages num
2026-01-04 11:03:16 +08:00
eead9538b2
[feat] update config
2025-12-31 10:29:13 +08:00
936d105e9b
[feat] compatible tokenizer
2025-12-31 10:26:46 +08:00
81e869fc3e
[fix] harmonize template
2025-11-06 13:14:08 +08:00
509d8dacf1
[feat] clear cache
2025-11-06 13:12:28 +08:00
8a0b04ed82
[fix] harmonize template
2025-11-02 23:18:11 +08:00
08ce3da228
[feat] update args
2025-10-30 10:48:31 +08:00
bf123b585d
[feat] add args
2025-10-30 10:05:12 +08:00
28cc44579a
[feat] convert2llama
2025-10-23 20:22:42 +08:00
7d877db79b
[feat] minimind-2510
2025-10-21 21:19:47 +08:00
2e118e9d3d
update convert script
2025-04-27 15:44:40 +08:00
a62faf34bd
250426
2025-04-26 10:05:47 +08:00
7da201a944
update chat-openai-api
2025-04-18 12:43:57 +08:00
4a758564e4
fix top_p float bug
2025-04-09 16:52:20 +08:00
08e9a22a25
update web_demo
2025-04-04 11:25:40 +08:00
c1a77f5c0f
update web_demo
2025-02-14 19:38:55 +08:00
d519d2a233
update web_demo
2025-02-14 19:37:27 +08:00
b5d10d9a7d
fix bugs
2025-02-13 20:56:14 +08:00
c9ca0f57dc
update config
2025-02-10 00:14:11 +08:00
29d6cd9cd2
update scripts file
2025-02-09 23:52:50 +08:00
58e3af0359
add minimind2
2025-02-09 23:49:47 +08:00
Justin Güse
tkgcc
jingyaogong
jingyaogong
gongjy