Views
No views yet
WAM_TEXT_EMBED_CACHE every backbone_type=wan22 / wan22_diffusers run in
Isaac-GR00T-AlinVLA-SH trains and evaluates against.| field | value |
|---|---|
| entries | 366 (365 raw tasks + the empty prompt) |
| embedding | (512, 4096) bf16, one per prompt |
| mask | (512,) bool, one per prompt |
projected | False — RAW umT5, not projected through any crossattn_proj |
max_sequence_length | 512 |
| encoder | Wan-AI/Wan2.2-TI2V-5B umT5 |
| tokenizer | Wan-AI/Wan2.1-T2V-1.3B |
formalize_language | True |
1export WAM_TEXT_EMBED_CACHE=/path/to/robocasa300_wan22_umt5.pt
2export WAM_TEXT_EMBED_REQUIRE_HIT=1 # umT5 encoder freed; every prompt must hit the cache'' must be present. Text dropout (WAM_TEXT_DROPOUT=0.1) and CFG feed
'', so with REQUIRE_HIT=1 a cache without it dies mid-run. This file has it. The training
scripts still check on startup and call
scripts/wam_dit4dit/append_empty_prompt_to_text_cache.py if it is missing.projected=False is load-bearing. These are raw umT5 states. A run whose crossattn_proj
was trained (the Cosmos joint lineage) needs live text encoding instead — a projected cache built
against a different proj silently mismatches. For the Wan lineage raw is correct: pair it with
WAN_TEXT_PAD_ZERO=1 and WAN_TEXT_ATTEND_ALL=1 so all 512 zero-padded positions are attended.scripts/wam_dit4dit/precompute_text_embeds_wan22.py.