• Joined on 2026-08-07
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:57:05 +00:00
9fbd6fe77e Batch-dim on input_ids for _embed_rows ([1, seq])
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:54:34 +00:00
0903b849b2 Flatten _text_run_ids output (input_ids is [1,seq])
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:51:51 +00:00
428c83d1d8 Coerce token ids to int in build_fl2va_presentation
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:48:27 +00:00
4add435bf9 Fix vision merger: main=per-patch-norm+2x2-interleave, deepstack=merged-norm
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:39:08 +00:00
68c49fdc95 Use merge_dim as normalized_shape in _VisionPatchMerger (4*1152 for 2x2 merge)
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:36:33 +00:00
b085ef02e8 Merge tokens expands C by merge^2 (merger takes pre-expanded)
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:34:18 +00:00
571d3541e8 Use explicit C (not x.shape[-1]) in _merge_tokens
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:31:55 +00:00
99a7550a51 Debug _merge_tokens prod
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:29:27 +00:00
ece7b2497e Use explicit C instead of -1 in _merge_tokens view
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:27:18 +00:00
834bd96111 Trace x shape at deepstack layers
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:21:31 +00:00
b9dcdda116 Debug _merge_tokens shapes
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:18:36 +00:00
ffb3c9e729 Debug print in vision tower forward
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:12:50 +00:00
e4d9a6f8ff Fix position_embeddings unpack (3-tuple of cos/sin/-sin)
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:10:41 +00:00
a19f051800 Cast qkv to common dtype before SDPA (q/k/v match)
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:08:24 +00:00
efc5fd9bc1 Fix col_idx expand/reshape precedence in rot_pos_emb
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:04:15 +00:00
67e4b12246 Make process_image return 2D grid_thw [[t,h,w]] for multi-image support
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 15:01:14 +00:00
6866263631 Fix F.conv3d call: remove invalid kernel_size kwarg (uses weight.shape)
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 14:56:34 +00:00
453aa86328 Use PIL Lanczos resize in resize_keyframe (F.interpolate no-4D-lanczos)
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 14:49:43 +00:00
0b7217485c nin_shortcut is a plain 1x1x1 conv (no causal padding)
daniel pushed to fl2va-direct-runtime at daniel/h3-blackwell-runtime 2026-08-19 14:47:02 +00:00
ea6ab87a34 Decouple spatial_padding from temporal_causal in causal conv