bingshuo/guanghulab
1
0
Fork 0
Code Issues 2 Pull Requests Actions 1 Packages Projects Releases Wiki Activity

training-status.json v1.2: OOM fixes in v7, distillation running stable

Browse Source
This commit is contained in:
bingshuo 2026-05-18 22:59:58 +08:00
parent 5a9c2db952
commit 93dea979e5
1 changed files with 11 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
brain/training-status.json
View File

@ -1,6 +1,6 @@
{
"version": "1.1",
"last_updated": "2026-05-18T21:00+08:00",
"version": "1.2",
"last_updated": "2026-05-18T22:45+08:00",
"updated_by": "铸渊 (ICE-GL-ZY001)",
"mother_model": {
"status": "completed",
@ -27,18 +27,19 @@
},
"distill_shuangyan_15b": {
"status": "running",
"status_label": "🔄 蒸馏训练中（PID 82143，step=50 loss=6.66，Epoch 1/3）",
"version": "v6",
"status_label": "🔄 蒸馏训练中（Screen 87779, step=50 loss=6.68, Epoch 1/3）",
"version": "v7",
"script": "scripts/distill_mother.py",
"teacher": "母模型7B SFT",
"student": "Qwen2.5-1.5B-Instruct",
"method": "KL散度(TEMP=2.0, ALPHA=0.7) + SFT混合",
"host": "connect.westd.seetacloud.com:23647",
"ssh_password": "HkM43lFVUIsc",
"started_at": "2026-05-18T20:56+08:00",
"estimated_completion": "~23:30+08:00",
"started_at": "2026-05-18T22:41+08:00",
"estimated_completion": "~01:00+08:00",
"cos_output": "models/shuangyan-15b-distill/final/",
"fixes": [
"v7: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True + empty_cache在epoch转换/模型加载/每500步",
"v6: vocab_size截断 fix — Teacher(7B)=152064 → Student(1.5B)=151936",
"v5: HuggingFace forward参数名修正 (ii→input_ids, am→attention_mask)",
"v4: num_workers=0 (避免CUDA fork)",
@ -52,20 +53,20 @@
{"name": "代码模型SFT训练", "status": "completed", "label": "✅ 已完成"},
{"name": "上传COS (代码模型)", "status": "completed", "label": "✅ 已完成"},
{"name": "generation_config修复", "status": "completed", "label": "✅ 两个模型已修复"},
{"name": "霜砚1.5B蒸馏", "status": "running", "label": "🔄 训练中（PID 82143, Epoch 1/3）"},
{"name": "霜砚1.5B蒸馏", "status": "running", "label": "🔄 训练中（Screen 87779, Epoch 1/3）"},
{"name": "上传COS (蒸馏输出)", "status": "pending", "label": "⚪ 等待蒸馏完成"}
]
},
"next_phases": {
"distill_shuangyan_15b": "🔄 训练中，预计~23:30完成",
"distill_shuangyan_15b": "🔄 训练中，预计~01:00完成",
"shuangyan_deep_sft": "⏳ 等待蒸馏完成后用霜砚语料深度微调",
"distill_coder_to_zhuyuan_15b": "⏳ 等待Phase 1+2完成",
"zhuyuan_deep_sft": "⏳ 等待Phase 3完成"
},
"alerts": [
"⚠️ 之前误判代码模型已完成（COS上有历史checkpoint-540）",
"⚠️ 蒸馏v1-v5多次崩溃：路径错误→存储爆炸→CUDA fork→参数名错误→vocab大小不匹配",
"⚠️ v6 vocab_size截断修复后训练中，预计~2.6小时完成"
"⚠️ 蒸馏经过7次迭代才稳定：路径错误→存储爆炸→CUDA fork→参数名错误→vocab不匹配→OOM(×2)",
"⚠️ v7 所有已知崩溃已修复，预计~01:00完成"
],
"links": {
"notion": "https://www.notion.so/363fb92f383181029fafef8438d6c300"