Commit Graph

22 Commits

Author SHA1 Message Date
21018fca8a fix: shared_experts missing ffn. prefix in checkpoint→model rename
Checkpoint keys are model.layers.N.shared_experts.gate_proj.weight
but model params are layers.N.ffn.shared_experts.gate_up_proj.weight.
The .ffn. was missing from the rename, so stacked gate_up_proj
never matched params_dict.
2026-05-15 00:17:59 +00:00
483046b9d6 fix: shared_experts gate_up_proj stacking was skipped by .experts. check
The stacking logic skipped any key containing '.experts.' to avoid
MoE routed expert weights. But 'shared_experts' also matches that
substring, so gate_proj and up_proj were never stacked into
gate_up_proj. Changed to '.ffn.experts.' which only matches the
routed experts path.

Also includes POST-LOAD all-zero param scan.
2026-05-15 00:08:04 +00:00
8dbd616add a little more debug1 2026-05-15 00:02:00 +00:00
756ea2192f clean up and possible big fix 2026-05-14 23:41:10 +00:00
9f01307c5b debug more7 2026-05-14 23:20:19 +00:00
e4f52c8900 debug more5 2026-05-14 23:01:59 +00:00
e46ff41569 debug more4 2026-05-14 22:50:51 +00:00
fd5f04eb15 debug more3 2026-05-14 22:36:34 +00:00
7573f12659 debug more2 2026-05-14 22:26:22 +00:00
11bbf675af debug more 2026-05-14 22:21:30 +00:00
5bbe51357c damn clankers 2026-05-14 20:23:42 +00:00
6aae8f1393 more fixes7 2026-05-14 20:11:37 +00:00
4363eee2ce more fixes6 2026-05-14 20:08:25 +00:00
40b980b9d6 more fixes5 2026-05-14 19:55:34 +00:00
d56e86b40e more fixes4 2026-05-14 19:51:56 +00:00
bf17bd3fc4 more fixes3 2026-05-14 19:47:02 +00:00
c68f4e9d6e more fixes2 2026-05-14 19:43:24 +00:00
4749a92fca more fixes 2026-05-14 19:39:16 +00:00
1ceff541b0 more fixes 2026-05-14 19:35:39 +00:00
3be051e140 fix 2026-05-14 19:29:47 +00:00
57512d5f0d clean up 2026-05-14 19:20:08 +00:00
0d8e1bd035 restructure: move Dockerfile and docker-compose to root, docker/ → vllm/ 2026-05-14 18:47:30 +00:00