ef89ceffbd
Add ModelOpt NVFP4 pipeline: patch, run script, README
...
- Patch fixes iter_weights_for_calibration() for DeepseekV4Experts
(ModuleList quantizers vs singular)
- Run script uses official NVIDIA hf_ptq.py with FP8 source
- Documents flags to avoid (--low_memory_mode, wrong arg names)
2026-05-07 07:22:54 +00:00
116933dcf6
Fix: skip .cuda() when low_memory_mode; switch default to nvfp4
2026-05-07 03:06:33 +00:00
b8bdd00d19
Lower GPU max_memory to 100GiB, add CPU-only fallback for low_memory_mode
2026-05-07 02:49:24 +00:00
717151b98c
Add CPU offloading and max_memory caps for FP8 model loading
2026-05-07 02:40:48 +00:00
aff12c6951
Fix forward_loop: pass as callable, not via create_forward_loop
2026-05-07 02:08:09 +00:00
492e44c0f6
Fix dataloader API: max_sample_length not seq_len, proper create_forward_loop
2026-05-07 02:04:54 +00:00
b32bb2e84d
NVIDIA Model Optimizer branch: nvfp4_experts_only PTQ for DeepSeek V4 Pro
2026-05-07 00:11:31 +00:00
c40607053b
Fix remaining gate_proj/up_proj -> w1/w3 references in paired_names
2026-05-07 00:05:55 +00:00
771e42cef3
Fix expert pair dict keys: w1/w3 not gate_proj/up_proj
2026-05-07 00:05:25 +00:00
5f35a5d2b3
Gracefully handle missing scale tensors (BF16 weights with stale index entries)
2026-05-07 00:04:29 +00:00
4470653e15
Fix V4 tensor naming: .scale companions, w1/w3 expert pairs, ffn.gate, hc_* preserve
2026-05-07 00:03:20 +00:00
2b7f063e39
7 commit
2026-05-06 23:51:54 +00:00
be16bd023e
sixth commit
2026-05-06 23:50:51 +00:00
97e7638abc
sixth commit
2026-05-06 23:49:34 +00:00
75503a1190
fifth commit
2026-05-06 23:49:02 +00:00
2eeeefcf8f
fourth commit
2026-05-06 23:48:38 +00:00
31a4302ab6
third commit
2026-05-06 23:48:25 +00:00
18ba8e057f
second commit
2026-05-06 23:47:38 +00:00
4708cdebb2
init commit
2026-05-06 23:47:07 +00:00
