Commit Graph

Select branches

Hide Pull Requests

4090-flex-compare

SolitaryThinker-patch-1

SolitaryThinker-patch-2

add_agents.md

add_ci

alex/opt-ltx23-nvfp4

aryan/h3-spark-gb10

attn-loud-fail

b200-ssim-full-quality

backend_dev

ci/precommit-status-post

ci/slash-test-precommit

codex/performance-dashboard-cohort-selection

codex/selectable-gpu-ci

comfy_custom_wendy

comfy_pipline

comp_time

cursor-agent

deltadit

dmd

docs-build

docs/add-release-skill

docs/readme-realtime-demo

ds8/dreamversev2-dev

experiments

feat/add-model-skill-revision

feat/cosmos3-action

feat/cosmos3-audio

feat/cosmos3-i2v

feat/cosmos3-reasoning

feat/cosmos3-tier-a-port

feat/dreamverse-monorepo-main-sync

feat/dreamverse-monorepo-main-sync-plus-profile

feat/h3-dmd2

feat/h3-dmd2-vsa

feat/h3-t2av-dmd-training

feat/improve-ci

feat/kaiqin/add-mg-df

feat/modular-training-profiler-qad

feat/pan2

feat/pipeline-startup-profiling

feat/startup-loader-foundation

feat/studio-ray-multinode

fix-h3-regional-compile-recompiles

fix/ci-retry-and-ssim

fix/fa4-cutedsl-api-compat

fix/interleave-thinker-review

fix/pr1807-collective-agreement

fix/setuptools-package-discovery

fix/vae-tiling-reset

fix/workload-type-from-registration

fix_timestep_sample

flex_sta_rlsu

flex_sta_rlsu_debug

generate_syn

hangliang_sw

hao-cleanup2

hao-edit

hao-test

hunyuan-gamecraft-v2

hunyuan15-support

hy_experiments

hy_experiments2

i2v_demo

i2v_fixed

impl-kandinsky-r12-repair

infer_sta_tea

inference-compile

infra/gb10-image-sm121a

integration/h3-perf-stack

integration/h3-vsa-fullgraph-all-20260822

jinzhe/distill

jinzhe/fix/parquet_dataset

jinzhe/flux

kernel/gb10-vsa128-tile128

kernels

klin/5090-fixes

klin/comfyui-precision-fixes

klin/dv-test

klin/h3-studio-editor

klin/trackwan

klin/ui-testing-1

klin/world_model_ui

kv_cache_fix

loay/ltx2_fp8

ltx25-integration

main

maint/ci-runner-rewrite-only

maint/cosmos3-final-refresh-20260720

maint/cosmos3-video-parity-20260720

maint/cosmos3-video-parity-rebase-only

maint/cosmos3-video-parity-seed

maint/dreamverse-remove-equal-delay-race

maint/ltx2-finetune-rebase-only

maint/ltx2-nvfp4-qat-rebase-only

maint/ltx2-rmsnorm-dtype

maint/perf-comparator-statuses

maint/perf-host-profile-guard

maint/perf-lane-cpu16

maint/perf-worker-log-capture

maint/pr1030-rebase-only

maint/pr1290-rebase-only

maint/pr1290-refresh

maint/pr1302-rebase-only

maint/pr1321-rebase-only

maint/pr1339-rebase-only

maint/pr1371-rebase-only

maint/pr1382-fullstack

maint/pr1382-rebase-only

maint/pr1442-fixed

maint/pr1442-rebase-only

maint/pr1449-fixed

maint/pr1453-fixed

maint/pr1453-rebase-only

maint/pr1461-rebase-only

maint/pr1471-fixed

maint/pr1471-fixed-v2

maint/pr1471-rebase-only

maint/pr1476-rebase-only

maint/pr1488-fixed

maint/pr1488-rebase-only

maint/pr1494-rebase-only

maint/pr1496-rebase-only

maint/pr1505-fixed

maint/pr1505-rebase-only

maint/pr1509-fixed

maint/pr1518-fixed

maint/pr1538-fixed

maint/pr1538-rebase-only

maint/pr1538-wave5

maint/pr1546-fixed

maint/pr1546-fixed-v2

maint/pr1551-fixed

maint/pr1560-rebase-only

maint/pr1561-rebase-only

maint/pr1563-rebase-only

maint/pr1574-rebase-only

maint/pr1577-rebase-only

maint/pr1579-rebase-only

maint/pr1580-rebase-only

maint/pr1581-rebase-only

maint/pr1590-rebase-only

maint/pr1591-rebase-only

maint/pr1596-rebase-only

maint/pr1601-rebase-only

maint/pr1602-rebase-only

maint/pr1617-torchcodec-fix

maint/pr1622-rebase-only

maint/pr1626-rebase-only

maint/pr1627-rebase-only

maint/pr1659-fixes

maint/pr1665-rebase-only

maint/pr1691-rebase-only

maint/pr1764-rebase-only

maint/pr1796-rebase-only

maint/pr1802-rebase-only

maint/preprocessing-ci

maint/skip-longcat-oom

maint/ssim-png-reference-cleanup

maint/zimage-ssim-prep

matthew/causal

matthew/debug_training

matthew/debug_training_2

matthew/demo

matthew/moe_debug

matthew/moe_revert

matthew/peprocessing

matthew/wan2.2-new

mochi-dev

model/port-lingbot-video-ti2v

model/port-magi-2

model_config

motion-stream/causal-kernel

ms-recipe

perf/fa4-vsa-block64

perf/fused-adamw-training

perf/h3-regional-compile-main-20260822

perf/h3-vsa-fullgraph-20260822

perf/ulysses-h3-adaptive

perf/ulysses-mfu-study

perf_reg_skill

pipeline_config

pr1362-fixes

pr1703-fixes

pr1711-fixes

py/add_rl

quant

readme-wan-crush-smol-fix

reason

refactor-config

refactor-lora

refactor/auto-merge-flow

refactor_attn

revert-1172-job_runner_ui_2

revert-591-yq/update

revert-595-main

revert-955-will/release

rlsu-test

rlsu-video-cut

satyam/ci-production-performance-regression

satyam/diagg_spark

satyam/profiling

satyam/temp_per_check_H3_spark

sdpa-mask-2d-fix

sdpa-mask-2d-fix-v2

shao/eval

shao/ltx2-cfg-v2

shreejith/unified-kernel-package

sta_do_not_delete

swe-1655-orphans

sync-branch

test-hf-sync

trackwan_bidir

upgrade_torch

v2

v2_backup

vae-parallel-decode

vsa_bwd_fix

wan-vae

wangame-debug-dfsft

wei/SF

wei/SF_infer

wei/api

wei/datasets

wei/distill

wei/dit_debug

wei/hy1.5

wei/hy15-1080

wei/hy15-i2v

wei/hy15_causal

wei/hy15_cf

wei/hy15_ode

wei/hy15_preprocess

wei/i2v

wei/issues

wei/ode_preprocess

wei/videox

wei/wan2.2-distill-new

wei/wan2.2-finetune

wei/wan2.2-training

wei/wan_dit_align

wei/wan_mc

will/add-model-skill

will/api_6_test

will/benchmark

will/cosmos2

will/debug

will/debug_ode

will/debug_vae

will/debug_validation

will/demo_public

will/design

will/distill_model_load

will/dmd_i2v_debug

will/docker_demo

will/dreamverse-monorepo

will/ema_fsdp_fix

will/f_0

will/fix-rebased

will/fix2

will/fix_denoise_init

will/fix_dep

will/fix_distill

will/fix_i2v

will/fix_shape

will/fix_tests

will/fix_validation

will/fix_validation_pipeline

will/fix_wan2

will/flux-ci

will/fp8

will/gradio-demo

will/hy15_training

will/i2v_oom

will/i2v_val

will/improve_preproc

will/improve_preprocessing

will/imrpove_api

will/kv_cache_fix_rebased

will/lora_checkpoint

will/ltx-base

will/ltx2

will/ltx2_sr

will/ltx2_sr_port

will/magi

will/magi-03-dit-review-fixes

will/magi-06-activate

will/moe_ode

will/move_tests

will/new_i2v

will/ode_init

will/preprocess_ci

will/process_prompts

will/qat

will/readme2

will/rebase

will/refactor_denoising

will/remove_v1_folder

will/reorg_examples

will/rewrite_validation

will/rl

will/sf_causal

will/sp_training

will/streaming

will/studio-l1-jobs

will/t2v_i2v

will/test_pr

will/training-VSA-rebase

will/training_mp

will/training_readme

will/transfusion

will/update_docker

will/update_min_python

will/v2.1

will/veo-training-data

will/videox

will/wan2.2

will/wan2.2-finetune-rebased-prof

will/wan2_2_sf

will/will_dmd_i2v

will/zmq-worker

william/pr1253-review-and-stable-audio

wip/flashinfer-sageattn3

worktree-agent-a58b1c80

worktree-agent-a67f7ec1

worktree-agent-a7083254

yq-update

yq/Wan2.2-time

yq/distill

yq/distill-i2v

yq/dmd-vis

yq/wan2.2-training

zihang-gradient-checkpoint

zihang-textencoder

zihang-training

v0.0.5

v0.1.0

v0.1.1

v0.1.2

v0.1.5

v0.1.6

v0.1.7

v0.2.0

Mono Color

• e95edb4218 [perf]: rebuild the deferred-release schedule when a stage is added late KyleandClaude Opus 5 2026-08-26 02:20:41 -07:00
• ed47a8a890 [perf]: expose lazy module loading in the FastH3 example KyleandClaude Opus 5 2026-08-25 17:52:34 -07:00
• d96dee46af [perf]: expose lazy module loading in the MiniMax H3 example KyleandClaude Opus 5 2026-08-25 15:17:00 -07:00
• 43644de268 [perf]: load pipeline components on demand and free them after their last stage KyleandClaude Opus 5 2026-08-25 14:27:29 -07:00
• a159b63c67 [bugfix] Harden OpenAI serving after post-merge review (#1782) William Lin 2026-08-28 22:29:24 -07:00
• ac48bb3cd1 [feat] Add MiniMax H3 MLX T2VA inference (#1770) 4 people 2026-08-28 12:54:33 -07:00
• 3987b9ddcd [feat] Align multimodal OpenAI serving APIs (#1781) William Lin 2026-08-28 10:09:58 -07:00
• c7da2f5d60 [chore]: release v0.2.1 (#1778) William Lin 2026-08-28 02:03:18 -07:00
• 39ae1decc0 [misc] pin fastvideo-kernel to exact 0.3.5 (#1777) William Lin 2026-08-28 02:02:27 -07:00
• 1aed667377 [chore] release fastvideo-kernel 0.3.5 (#1775) William Lin 2026-08-27 23:18:59 -07:00
• c1612ff397 [bugfix]: pin fastvideo-kernel to Torch 2.12.0 (#1774) William Lin 2026-08-27 23:13:54 -07:00
• 5a06b460b2 Fix Ulysses sequence-parallel crash in self-forcing causal Wan DiT maint/pr1627-rebase-only Peter KisfaludiandClaude Opus 4.8 2026-07-20 18:04:16 -07:00
• a534ba20a0 [feat] Add MiniMax H3 LoRA inference and preview launchers (#1771) William Linandshaoxiongduan 2026-08-27 14:43:39 -07:00
• be1df9fec5 [ci][diagnostic]: capture GameCraft FA4 candidate on GB200 maint/pr1764-rebase-only William Lin 2026-08-26 12:14:19 -07:00
• e9bbaca07d [perf] Disable every offload path on unified memory, unblocking MiniMax H3 generation on one GB10 (#1715) KyleNeverGivesUp 2026-08-26 15:32:56 -07:00
• 9bfa585448 [perf]: stop holding the whole checkpoint during DiT load, unblocking MiniMax H3 on one GB10 (#1714) KyleNeverGivesUp 2026-08-26 15:23:51 -07:00
• b2062556a9 [perf] VSA Triton: widen the autotune num_stages range (the optimum was outside it) (#1706) Raghav K 2026-08-26 12:30:56 -07:00
• c9c5585758 [perf]: MiniMax H3 on GB10 - skip text encoder CPU offload on unified memory (5m49s to 30ms) (#1710) KyleNeverGivesUp 2026-08-26 12:03:14 -07:00
• 9212f4f218 [ci] make GPU validation change-aware (#1747) William Lin 2026-08-25 21:26:25 -07:00
• 6388db815b [bugfix] FastMetal-QAD MLX support: refuse CUDA QAD trees, use packed mlx_dit config, stream loads (#1736) (#1758) Aryan KumarandAryan Kumar 2026-08-25 14:51:40 -07:00
• 7a4285189f [kernel] Route block-sparse VSA to the sm_100a forward behind FASTVIDEO_VSA_SM100A (opt-in) (#1754) lpc0220 2026-08-24 15:26:56 -07:00
• a837fe841a [docs] Update FastH3 README (#1749) William Lin 2026-08-23 07:05:58 -05:00
• f9e3680f11 [perf] Align FastH3 optimized inference profile (#1748) William Lin 2026-08-23 04:01:03 -05:00
• 98f761ec45 [bugfix] validation: inherit the trained denoising ladder (#1738) William Lin 2026-08-23 01:09:26 -05:00
• c041318f2c [perf] Add fused NVLink all-to-all for Ulysses (#1740) Shao Duan 2026-08-23 14:06:24 +08:00
• 604e0205a4 [perf] Keep odd MiniMax-H3 VSA tiles on sm100a (#1745) William Lin 2026-08-22 20:29:30 -05:00
• 13213395b4 [perf] Parallelize MiniMax-H3 VAE over sequence ranks (#1744) William Lin 2026-08-22 20:05:31 -05:00
• 46afee5998 [bugfix] Classify MiniMax-H3 inference controls in schema inventory (#1743) William Lin 2026-08-22 19:17:02 -05:00
• c488fa1211 [perf] Add opt-in packed-varlen FA4 for MiniMax-H3 (#1742) William Lin 2026-08-22 19:16:47 -05:00
• d3cff517cd [perf] Add opt-in regional fullgraph compile for DiT inference (#1741) William Lin 2026-08-22 14:14:39 -05:00
• 6c51129ad0 [test] MiniMax-H3: use supported graph-reuse geometry integration/h3-vsa-fullgraph-all-20260822 Will Lin 2026-08-22 06:39:45 +00:00
• fae4f82d6e [test] MiniMax-H3: use supported graph-reuse geometry perf/h3-vsa-fullgraph-20260822 Will Lin 2026-08-22 06:39:45 +00:00
• d137c51919 [bugfix] MiniMax-H3: prepare compile after FP8 conversion Will Lin 2026-08-22 06:38:09 +00:00
• d54f694339 [bugfix] MiniMax-H3: prepare compile after FP8 conversion Will Lin 2026-08-22 06:38:09 +00:00
• 37927079ff [perf] MiniMax-H3: share compiled VSA graph across layers Will Lin 2026-08-22 06:35:55 +00:00
• 75c41e7e35 [perf] MiniMax-H3: share compiled VSA graph across layers Will Lin 2026-08-22 06:35:55 +00:00
• 9432a87ee5 [perf] MiniMax-H3: compile VSA attention regions Will Lin 2026-08-22 06:19:42 +00:00
• 658c56d001 [perf] kernel: make sm100a sparse inference compile-safe Will Lin 2026-08-22 06:19:37 +00:00
• d20484ff97 [perf] MiniMax-H3: compile VSA attention regions Will Lin 2026-08-22 06:19:42 +00:00
• c4db339249 [perf] kernel: make sm100a sparse inference compile-safe Will Lin 2026-08-22 06:19:37 +00:00
• 99cd355a24 [perf]: use packed-varlen FA4 for MiniMax H3 Will Lin 2026-08-22 04:25:54 +00:00
• 172d1b5dfc [docs] integrate final parallel H3 VAE parity caveat Will Lin 2026-08-22 03:38:24 +00:00
• 35cf5a5096 [perf] integrate scoped regional H3 compile policy Will Lin 2026-08-22 03:38:23 +00:00
• 8518f36ec1 Merge commit 'f1e3cf751' into perf/h3-compile-safe-fusions-20260822 Will Lin 2026-08-22 03:36:01 +00:00
• f1e3cf7511 [bugfix]: scope regional attention compile opt-in perf/h3-regional-compile-main-20260822 Will Lin 2026-08-22 03:35:05 +00:00
• 27f499a6a8 [docs] clarify compiled H3 VAE parity envelope Will Lin 2026-08-22 03:27:10 +00:00
• 3a8f463d44 [perf] integrate final compiled H3 fusion messaging Will Lin 2026-08-22 03:17:09 +00:00
• dc82dbc3e6 [docs] MiniMax-H3: describe compiled fusion boundaries accurately Will Lin 2026-08-22 03:17:05 +00:00
• da572d8584 [test] integrate hardened compiled H3 fusion gates Will Lin 2026-08-22 03:14:17 +00:00
• 279071df7e [test] gate compiled H3 fusion dispatcher contracts Will Lin 2026-08-22 03:13:11 +00:00
• 84766e0174 [perf] integrate compile-safe H3 fusions for matched benchmarking Will Lin 2026-08-22 03:09:42 +00:00
• 9709cc037e [perf] MiniMax-H3: keep Triton fusions active under regional compile Will Lin 2026-08-22 03:09:14 +00:00
• a72cfddade [perf] integrate H3 regional compile and parallel VAE for benchmarking Will Lin 2026-08-22 02:58:42 +00:00
• 8f57088128 [docs] H3 parallel VAE: document the compiled-decoder cross-process determinism caveat Will LinandClaude Fable 5 2026-08-21 18:17:28 +00:00
• 2875831e21 [perf] H3 parallel VAE: overlap first decodes with the meta rendezvous; assembly on a side stream Will LinandClaude Fable 5 2026-08-21 17:59:18 +00:00
• daac1ace14 [docs] schema parity inventory: classify vae_parallel_* (+ the stack's unclassified VSA_tile_size) Will LinandClaude Fable 5 2026-08-21 17:53:41 +00:00
• 2e88101230 [perf] MiniMax-H3 VAE: SP-rank-parallel chunk decode + reference-clip encode (opt-in) Will LinandClaude Fable 5 2026-08-21 17:38:38 +00:00
• 2a9549166b [bugfix] logger: info_once crashed on the patched process-aware info (duplicate stacklevel) Will LinandClaude Fable 5 2026-08-21 17:38:18 +00:00
• 72759cb7de [docs] inference optimizations: document the regional fullgraph compile knob Will LinandClaude Fable 5 2026-08-21 17:55:03 +00:00
• 129fcaffab [perf] inference: regional fullgraph torch.compile of the DiT blocks (port of #1718 to the inference path) Will LinandClaude Fable 5 2026-08-21 17:46:26 +00:00
• 2f3d407406 [perf] Optimize MiniMax H3 VAE decoding (#1734) Junda Su 2026-08-21 14:57:32 -07:00
• bcffa4026e [perf] Optimize MiniMax-H3 text encoder memory (#1732) Kaiqin Kong 2026-08-21 14:57:06 -07:00
• 0a861031c7 [docs] H3 parallel VAE: document the compiled-decoder cross-process determinism caveat vae-parallel-decode Will LinandClaude Fable 5 2026-08-21 18:17:28 +00:00
• f08c5ee8af [perf] H3 parallel VAE: overlap first decodes with the meta rendezvous; assembly on a side stream Will LinandClaude Fable 5 2026-08-21 17:59:18 +00:00
• 9ba740e125 [docs] inference optimizations: document the regional fullgraph compile knob inference-compile Will LinandClaude Fable 5 2026-08-21 17:55:03 +00:00
• 755f4a7967 [docs] schema parity inventory: classify vae_parallel_* (+ the stack's unclassified VSA_tile_size) Will LinandClaude Fable 5 2026-08-21 17:53:41 +00:00
• 34346779d1 [perf] inference: regional fullgraph torch.compile of the DiT blocks (port of #1718 to the inference path) Will LinandClaude Fable 5 2026-08-21 17:46:26 +00:00
• 6d6a10be7a [feat] FastVideo-Minimax-FastH3-Preview few-step example + 64-token-tile VSA-H3 inference path (#1731) William Lin 2026-08-21 12:40:09 -05:00
• 73dd105f3d [perf] Add opt-in MiniMax-H3 Sol-Engine fusions (#1735) Kaiqin Kong 2026-08-21 10:39:28 -07:00
• d543a67b10 [perf] MiniMax-H3 VAE: SP-rank-parallel chunk decode + reference-clip encode (opt-in) Will LinandClaude Fable 5 2026-08-21 17:38:38 +00:00
• 741aa8d289 [bugfix] logger: info_once crashed on the patched process-aware info (duplicate stacklevel) Will LinandClaude Fable 5 2026-08-21 17:38:18 +00:00
• ad9cd63122 [bugfix] H3 conditioner: no_grad instead of inference_mode - FSDP2-sharded encode crashed (#1732 default-path gap) integration/h3-perf-stack Will LinandClaude Fable 5 2026-08-21 10:49:47 +00:00
• 68e6ffca9e [bugfix] H3 VAE: clone the reduce-overhead stitched canvas at the tile-driver returns (#1734 review F1) Will LinandClaude Fable 5 2026-08-21 10:33:06 +00:00
• 64cdcf6be4 Merge PR #1731 head (9713ea127) on top of the stack (the target of this branch) Will Lin 2026-08-21 10:26:31 +00:00
• aa0d98a6b8 Merge PR #1734 head (dca423fd3) into integration/h3-perf-stack Will Lin 2026-08-21 10:25:17 +00:00
• 93b03bc14d Merge PR #1735 head (cbab605ef) into integration/h3-perf-stack Will Lin 2026-08-21 10:22:36 +00:00
• 160f0c9ccf Merge PR #1732 head (ac56806af) into integration/h3-perf-stack Will Lin 2026-08-21 10:22:10 +00:00
• e5d1110a0f Merge PR #1703 head (pr1703-fixes @ 942f7db3d) into integration/h3-perf-stack Will Lin 2026-08-21 10:20:55 +00:00
• 622217ff2a Merge PR #1362 head (pr1362-fixes @ aa95a4c18) into integration/h3-perf-stack Will Lin 2026-08-21 10:20:46 +00:00
• cbab605eff [misc] MiniMax-H3 fusions: exact eager fallback, enable/inert logging, engagement test Will LinandClaude Fable 5 2026-08-21 10:06:23 +00:00
• ac98869aa1 [bugfix] MiniMax-H3 fusions: int64 row offsets in the fused qknorm+RoPE kernel Will LinandClaude Fable 5 2026-08-21 10:06:12 +00:00
• 56d4a6074f [bugfix] fastvideo-kernel: fix Triton block-sparse backward logit scaling (bf16 K pre-scaling) (#1730) William LinandClaude Fable 5 2026-08-21 04:50:06 -05:00
• 9713ea1275 [misc] use the full release name FastVideo-Minimax-FastH3-Preview-v0.1 Will LinandClaude Fable 5 2026-08-21 05:29:10 +00:00
• 37aa382cce [bugfix] example: the distilled FastH3 grid is 4 forwards = a 5-point sigma grid Will LinandClaude Fable 5 2026-08-21 04:42:38 +00:00
• 3f00983287 [feat] attention: opt-in sm_100a CUDA forward route for VSA-H3 tile-64 Will LinandClaude Fable 5 2026-08-21 02:49:17 +00:00
• 2dc57f4070 [misc] rebrand the few-step preview to FastH3 (model string, example name, docs) Will LinandClaude Fable 5 2026-08-21 02:58:22 +00:00
• 9df19be719 [feat] example: few-step FastVideo-Minimax-H3-Preview inference (4-step DMD2 student) Will LinandClaude Fable 5 2026-08-21 02:43:30 +00:00
• 089eea3970 [feat] inference: expose the VSA-H3 tile size on the run-level route Will LinandClaude Fable 5 2026-08-21 02:36:11 +00:00
• dd8447ecc5 [feat] VSA-H3: 64-token tile option on the native Triton block-sparse path Will LinandClaude Fable 5 2026-08-21 02:34:06 +00:00
• dca423fd31 Enable H3 VAE attention backend selection. Davids048 2026-08-21 06:46:00 +00:00
• 1b43af8e8e Custom compile H3 vae decoding. Davids048 2026-08-21 05:52:59 +00:00
• 628591b620 Add some logging. Davids048 2026-08-21 05:18:44 +00:00
• 0980ca563f Add nvtx profiling support. Davids048 2026-08-21 05:14:48 +00:00
• b158388733 [perf] Add opt-in MiniMax-H3 Sol-Engine fusions H1yori233 2026-08-21 00:18:56 -07:00
• c4ad4227c0 [misc] MiniMax-H3: move the AdaLN converter into scripts/checkpoint_conversion (#1712) Shao Duan 2026-08-21 15:15:00 +08:00
• a63ccce73d [docs]: add a maintained inference cookbook (#1290) KyleNeverGivesUpandSolitaryThinker 2026-08-20 23:53:44 -07:00
• ac56806aff [perf] Optimize MiniMax-H3 text encoder H1yori233andWill Lin 2026-08-20 20:42:10 -07:00
• 0462e1b0e7 [perf]: MiniMax H3 - build the Qwen3-VL encoder only as far as it is read (-13.7 GB) (#1711) KyleNeverGivesUp 2026-08-20 21:19:25 -07:00
• 907f2100ec [kernel] sm_100a CUDA block-sparse VSA forward (Blackwell), 64- and 128-token blocks (#1719) lpc0220 2026-08-20 21:16:25 -07:00
• e0a3db5651 [perf] Reduce MiniMax-H3 VAE peak memory (#1703) Kaiqin Kong 2026-08-20 21:13:23 -07:00