h3-blackwell-runtime

daniel/h3-blackwell-runtime

Fork 0

Commit graph

d345aae6e9 Record post-FC2 production profile fl2va-direct-runtime Daniel Maddern 2026-08-26 16:14:12 +07:00
67e05f517c Validate aligned Vortex attention prototype Daniel Maddern 2026-08-26 15:51:39 +07:00
c58e4c6d12 Validate Vortex attention capabilities Daniel Maddern 2026-08-26 14:44:28 +07:00
ff8f7672b3 Complete Vortex attention interface design Daniel Maddern 2026-08-26 13:40:59 +07:00
7d0608e857 Register Vortex exact attention research Daniel Maddern 2026-08-26 13:40:16 +07:00
b842badfc1 Select Vortex exact attention architecture Daniel Maddern 2026-08-26 13:39:20 +07:00
c5028427d7 Define Vortex exact attention contract Daniel Maddern 2026-08-26 13:20:33 +07:00
a29b8960b0 Integrate guarded FC2 cuBLASLt schedule Daniel Maddern 2026-08-26 00:48:44 +07:00
0298ccd1cb Record exact FC2 cuBLASLt schedule Daniel Maddern 2026-08-25 22:32:48 +07:00
ea14b98544 Preserve Vortex performance research history Daniel Maddern 2026-08-25 20:30:22 +07:00
bee998d840 Record RTX Pro 6000 multi-GPU scaling master Daniel Maddern 2026-08-22 16:42:09 +07:00
6d8a1c3c2b Record RTX Pro 6000 SDPA baseline Daniel Maddern 2026-08-22 15:38:09 +07:00
49a9d0d1a1 Fix RunPod API client requests Daniel Maddern 2026-08-22 14:13:15 +07:00
bd92baeb46 Add distributed H3 execution and validation Daniel Maddern 2026-08-22 14:09:45 +07:00
1d5faa8f16 Expand H3 runtime acceleration and validation Daniel Maddern 2026-08-20 22:08:52 +07:00
837c16871d Add resident FL2VA Turbo variants Daniel Maddern 2026-08-20 19:13:23 +07:00
6d8c9ca4cf Add selectable hot attention backends Daniel Maddern 2026-08-20 17:39:44 +07:00
9bb96a26e8 Fix FL2VA conditioning parity end to end Daniel Maddern 2026-08-20 16:43:22 +07:00
9f62f6ea83 Remove debug prints from vision tower (fl2va passes e2e smoke) Daniel Maddern 2026-08-19 23:03:58 +07:00
45d3e7bc64 Fix mrope_freqs_cis call (no device kwarg) Daniel Maddern 2026-08-19 22:59:26 +07:00
9fbd6fe77e Batch-dim on input_ids for _embed_rows ([1, seq]) Daniel Maddern 2026-08-19 22:57:03 +07:00
0903b849b2 Flatten _text_run_ids output (input_ids is [1,seq]) Daniel Maddern 2026-08-19 22:54:32 +07:00
428c83d1d8 Coerce token ids to int in build_fl2va_presentation Daniel Maddern 2026-08-19 22:51:48 +07:00
4add435bf9 Fix vision merger: main=per-patch-norm+2x2-interleave, deepstack=merged-norm Daniel Maddern 2026-08-19 22:48:26 +07:00
68c49fdc95 Use merge_dim as normalized_shape in _VisionPatchMerger (4*1152 for 2x2 merge) Daniel Maddern 2026-08-19 22:39:05 +07:00
b085ef02e8 Merge tokens expands C by merge^2 (merger takes pre-expanded) Daniel Maddern 2026-08-19 22:36:32 +07:00
571d3541e8 Use explicit C (not x.shape[-1]) in _merge_tokens Daniel Maddern 2026-08-19 22:34:16 +07:00
99a7550a51 Debug _merge_tokens prod Daniel Maddern 2026-08-19 22:31:52 +07:00
ece7b2497e Use explicit C instead of -1 in _merge_tokens view Daniel Maddern 2026-08-19 22:29:26 +07:00
834bd96111 Trace x shape at deepstack layers Daniel Maddern 2026-08-19 22:27:16 +07:00
b9dcdda116 Debug _merge_tokens shapes Daniel Maddern 2026-08-19 22:21:30 +07:00
ffb3c9e729 Debug print in vision tower forward Daniel Maddern 2026-08-19 22:18:33 +07:00
e4d9a6f8ff Fix position_embeddings unpack (3-tuple of cos/sin/-sin) Daniel Maddern 2026-08-19 22:12:48 +07:00
a19f051800 Cast qkv to common dtype before SDPA (q/k/v match) Daniel Maddern 2026-08-19 22:10:39 +07:00
efc5fd9bc1 Fix col_idx expand/reshape precedence in rot_pos_emb Daniel Maddern 2026-08-19 22:08:21 +07:00
67e4b12246 Make process_image return 2D grid_thw [[t,h,w]] for multi-image support Daniel Maddern 2026-08-19 22:04:12 +07:00
6866263631 Fix F.conv3d call: remove invalid kernel_size kwarg (uses weight.shape) Daniel Maddern 2026-08-19 22:01:12 +07:00
453aa86328 Use PIL Lanczos resize in resize_keyframe (F.interpolate no-4D-lanczos) Daniel Maddern 2026-08-19 21:56:32 +07:00
0b7217485c nin_shortcut is a plain 1x1x1 conv (no causal padding) Daniel Maddern 2026-08-19 21:49:40 +07:00
ea6ab87a34 Decouple spatial_padding from temporal_causal in causal conv Daniel Maddern 2026-08-19 21:47:00 +07:00
b557cba173 Fix Downsample3D spatial reflect pad to (W,H) dims only Daniel Maddern 2026-08-19 21:41:16 +07:00
136297f4ae Fix F.pad dim order (spatial reflect + no T pad) for 1-frame path Daniel Maddern 2026-08-19 21:36:40 +07:00
c2d507691b Restore front-zero temporal pad for multi-frame causal conv; two-step F.pad (reflect spatial + constant T) Daniel Maddern 2026-08-19 21:35:07 +07:00
f7c48b067a Remove permute bug: rely on F.conv3d spatial padding Daniel Maddern 2026-08-19 21:33:08 +07:00
6fc7cf0ad6 Fix causal conv: correct F.pad spatial dim order + 1-frame kernel truncation Daniel Maddern 2026-08-19 21:31:05 +07:00
3386975326 Fix VAE: use same multi-frame causal path for keyframes, add frame_pre_padding so 1 keyframe -> 1 latent Daniel Maddern 2026-08-19 21:20:32 +07:00
a0274a9868 Fix tiled_encode latent x-overlap off-by-one (use [j-1] for left neighbor) Daniel Maddern 2026-08-19 21:13:40 +07:00
65e80be1cf Thread single_frame flag so keyframe truncation does not apply to 1-frame tiles Daniel Maddern 2026-08-19 21:11:43 +07:00
06fd79a8fd Fix causal temporal padding to match reference (2k-1 front zeros when spatial_padding>0) Daniel Maddern 2026-08-19 21:09:03 +07:00
390135fca6 Fix VAE pixel normalization in-place bug Daniel Maddern 2026-08-19 21:07:34 +07:00
9bec53ac36 Fix VAE encoder to load canonical checkpoint key names Daniel Maddern 2026-08-19 21:04:30 +07:00
807bd64a82 Add direct first/last-frame (fl2va) keyframe conditioning Daniel Maddern 2026-08-19 20:17:41 +07:00
8730920634 Default Spark runs to Sol attention vae-decode-optimization Daniel Maddern 2026-08-15 03:37:50 +07:00
731a2813fd Add Sol native attention baseline Daniel Maddern 2026-08-15 03:35:59 +07:00
c2fb5c1e3f Document NVFP4 native pack status Daniel Maddern 2026-08-15 03:00:06 +07:00
8d90cec133 Fix pack profiler imports Daniel Maddern 2026-08-15 02:57:21 +07:00
ddc0c5a2b7 Add NVFP4 pack geometry profiler Daniel Maddern 2026-08-15 02:56:19 +07:00
d4c243c9ef Match NVFP4 tie-to-even rounding Daniel Maddern 2026-08-15 02:48:06 +07:00
11d3c4164e Match NVFP4 midpoint rounding Daniel Maddern 2026-08-15 02:45:16 +07:00
9c0992350f Match NVFP4 signed zero encoding Daniel Maddern 2026-08-15 02:43:03 +07:00
ef84a2786c Match H3 NVFP4 row scale swizzle Daniel Maddern 2026-08-15 02:40:09 +07:00
fbfb3453de Zero initialize native NVFP4 block scales Daniel Maddern 2026-08-15 02:38:08 +07:00
7f9b1560e4 Add H3 NVFP4 scale column swizzles Daniel Maddern 2026-08-15 02:34:52 +07:00
1200152d79 Match wide NVFP4 block scale swizzle Daniel Maddern 2026-08-15 02:30:41 +07:00
b3a2544936 Match NVFP4 block scale swizzle Daniel Maddern 2026-08-15 02:26:48 +07:00
57b84b8d6b Match NVFP4 block scale padding Daniel Maddern 2026-08-15 02:24:09 +07:00
0e855b8e55 Prototype Vortex native NVFP4 quantizer Daniel Maddern 2026-08-15 02:22:37 +07:00
1fec77d2e9 Tune Vortex NVFP4 scale geometry Daniel Maddern 2026-08-15 02:15:19 +07:00
1cc0ffdd00 Add Vortex NVFP4 scale v1 Daniel Maddern 2026-08-15 02:11:47 +07:00
684b645546 Match Torch NVFP4 scale precision Daniel Maddern 2026-08-15 01:57:25 +07:00
0fda84502a Add Vortex NVFP4 scale kernel Daniel Maddern 2026-08-15 01:54:02 +07:00
03dbe1456d Prototype Vortex NVFP4 quantizer seam Daniel Maddern 2026-08-15 01:47:03 +07:00
d6eabb150e Add NVFP4 linear profiler Daniel Maddern 2026-08-15 01:37:42 +07:00
7cc03f3e57 Add H3 block profiler plan Daniel Maddern 2026-08-15 01:32:18 +07:00
3d0c093168 Add Sol-Attn backend Daniel Maddern 2026-08-14 21:04:32 +07:00
75ee9ba4ca Document backend implementation status Daniel Maddern 2026-08-14 20:39:20 +07:00
c3cc04e98d Add approximate H3 cache modes Daniel Maddern 2026-08-14 20:38:48 +07:00
fb257ef982 Add exact memory backend options Daniel Maddern 2026-08-14 20:34:41 +07:00
eaf9324145 Add KJ Sage attention backends Daniel Maddern 2026-08-14 20:28:56 +07:00
1c0883a54b Add persistent hot runtime service Daniel Maddern 2026-08-14 20:13:48 +07:00
54dd649ccf Plan persistent hot runtime Daniel Maddern 2026-08-14 20:02:47 +07:00
d1cdd4b6b3 Record Sage3 cat benchmark Daniel Maddern 2026-08-14 20:02:14 +07:00
a388aa64ff Default video VAE decode to FP16 Daniel Maddern 2026-08-14 19:50:36 +07:00
ea7712b4b2 Use paired kitchen RoPE in VAE fast path Daniel Maddern 2026-08-14 14:58:11 +07:00
8595dd875e Add optional kitchen VAE fast ops Daniel Maddern 2026-08-14 14:54:38 +07:00
f6166ede68 Add VAE decode benchmark tool Daniel Maddern 2026-08-14 14:36:42 +07:00
ea4f4a2dad Add Comfy cat benchmark submitter Daniel Maddern 2026-08-14 14:20:45 +07:00
53cd8bd4f4 Make runtime diagnostics opt in Daniel Maddern 2026-08-14 14:11:36 +07:00
fea69673cd Return native H3 audio latents Daniel Maddern 2026-08-14 13:57:30 +07:00
59b8302d4e Add direct H3 audio VAE decode Daniel Maddern 2026-08-14 13:39:46 +07:00
c5cf42850b Document H3 audio VAE path Daniel Maddern 2026-08-14 13:33:08 +07:00
4696a3e72c Add safetensors shape inspector Daniel Maddern 2026-08-14 13:32:01 +07:00
b73d08b231 Compute upstream VAE temporal overlap constants Daniel Maddern 2026-08-14 13:04:58 +07:00
81714ed50f Trace VAE temporal blend metadata Daniel Maddern 2026-08-14 12:56:36 +07:00
26fe7be966 Document VAE debug handoff Daniel Maddern 2026-08-14 01:58:27 +07:00
feeb4b75c9 Trace VAE blend inputs Daniel Maddern 2026-08-14 01:50:14 +07:00
28a4121df9 Match upstream VAE blend implementation Daniel Maddern 2026-08-14 01:41:53 +07:00
4e17e866a0 Add VAE temporal assembly tracer Daniel Maddern 2026-08-14 01:33:33 +07:00
49bae4ab74 Match upstream VAE temporal padding Daniel Maddern 2026-08-14 01:25:53 +07:00
6dd0b0e737 Add VAE tiled clip comparator Daniel Maddern 2026-08-14 01:20:27 +07:00