Commit Graph
100 Commits
Author SHA1 Message Date
MickandClaude Opus 5 ab9750fb35 [diffusion] optimization: pin layerwise host stores in place at their exact size (#39021)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-09-11 22:04:40 +08:00
MickandMick Qian 358c163250 [diffusion] fix: recover ipc jit initialization after interrupted builds (#39034)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-09-11 19:34:09 +08:00
MickandMick Qian c9c26d56b2 [diffusion] docs: sync snapshot and minimax-h3 subblock features (#38784)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-09-10 17:23:43 +08:00
Mick 7f0285093e [diffusion] feat: spill large tensors over shared memory like numpy arrays (#38656) 2026-09-10 09:14:06 +08:00
MickandMick Qian ce555ed82a [diffusion] refactor: refactor utility ownership and document helper placement (#38699)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-09-10 09:11:01 +08:00
MickandMick Qian 00a9028e87 [diffusion] feat: add explicit snapshot-offload component residency (#38535)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-09-09 13:44:54 +08:00
MickandMick Qian 0ee8e41a4e [diffusion] CI: don't fail the nightly when a perf dump is unreadable (#38629)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-09-09 13:39:10 +08:00
MickandMick Qian a8e45f16cc [diffusion] feat: support mixed INT8 embeddings and Comfy NVFP4 encoders for minimax-h3 (#38506)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-09-09 08:36:43 +08:00
MickandMick Qian 65400bb420 [diffusion] model: support MiniMax-H3 singularity hybrid checkpoints (#38455)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-09-09 08:35:39 +08:00
15ff470472 [diffusion] chore: key the VAE decode-dtype store by module layout (#38496)
Co-authored-by: Mick Qian <mickqian@radixark.ai>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-09-08 22:37:21 +08:00
88a9bfd1ff [diffusion] CI: add per-case timing tolerance for host-I/O-bound perf guards (#38457)
Co-authored-by: Mick Qian <mickqian@radixark.ai>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-09-08 18:40:17 +08:00
775f17b07c [diffusion] optimization: stream mapped weights on a shared host/device pool (#38441)
Co-authored-by: Mick Qian <mickqian@radixark.ai>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-08 17:52:52 +08:00
f3ccd1c0e4 [diffusion] CI: baseline the e2e of ten unguarded perf cases (#38335)
Co-authored-by: Mick Qian <mickqian@radixark.ai>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-07 23:13:14 +08:00
Mick ba6d3df69a [diffusion] doc: document verified GB300 and derived GB200 H3 recipes (#38296) 2026-09-07 16:32:29 +08:00
Mick 15d2cbcc90 [diffusion] CI: validate every repeated server request (#38185) 2026-09-07 14:36:50 +08:00
Mick ff08bcdda9 [diffusion] UX: quiet internal warmup frame searches (#38226) 2026-09-07 09:49:13 +08:00
Mick b83f1bdd21 [diffusion] fix: stabilize H3 reference audio across repeated requests (#38225) 2026-09-07 09:32:57 +08:00
Mick f3d05644db [diffusion] docs+skill: document which components to stream under layerwise offload (#35674) 2026-09-06 23:12:36 +08:00
MickandClaude Fable 5 ade1da017f [diffusion] docs: verify the DGX Spark H3 recipe (#37456)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-09-06 23:10:27 +08:00
MickandClaude Fable 5.1 a176ba2f7b [diffusion] feat: measure warmup memory and layer usage per phase for residency calibration (1/4) (#37916)
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-06 19:42:42 +08:00
Mick 938dc5621d [diffusion] refactor: reuse plain state-dict loading without per-model classes (#38127) 2026-09-06 18:39:21 +08:00
Mick a9944aec01 [diffusion] CI: guard the allocated vram peak with reporting the reserved one (#38172) 2026-09-06 17:22:16 +08:00
Mickandmickqian 8ef646a5c6 fix(vlm): contain EPD request lifecycle failures (#36944)
Co-authored-by: mickqian <mickqian@users.noreply.github.com>
2026-09-06 16:05:10 +08:00
Mick e3f7097591 [diffusion] refactor: consolidate plain state-dict component loaders (#38128) 2026-09-06 13:52:54 +08:00
Mick febb360519 [VLM] retire aborted disaggregated prefill results (#36988) 2026-09-06 10:15:33 +08:00
Mick 5df60a21cd fix(vlm): harden EPD receiver validation and liveness (#36945) 2026-09-05 21:22:37 +08:00
Mick a18106bbc3 fix(vlm): make EPD cache publication transactional (#36949) 2026-09-05 20:33:23 +08:00
Mick 0ea8378085 [diffusion] feat: support request-scoped skip-softmax attention (#37959) 2026-09-05 13:50:17 +08:00
Mick 88021b0734 [diffusion] chore: make nightly performance measurements robust (#37915) 2026-09-04 22:03:25 +08:00
Mick ed67b73081 [diffusion] UX: reduce hot-path server log noise (#37804) 2026-09-04 13:31:00 +08:00
Mick 97d081ac76 [diffusion] chore: remove unreachable cosmos3 transfer encoding (#37805) 2026-09-04 12:14:03 +08:00
Mick 2a0602c7ac [diffusion] optimization: unfused w2 bias on SM12.x (cuBLAS 16x16 kernel mis-dispatch) for minimax-h3 vae decoder: (#37835) 2026-09-04 09:17:23 +08:00
Mick 795dd7abce [diffusion] feat: compose third-party component bundles safely (#37816) 2026-09-04 09:15:05 +08:00
Mick 0dd66def7c [chore] harden checkpoint quantization metadata parsing (#36922) 2026-09-03 09:27:22 +08:00
Mick f6aed6ec53 [diffusion] doc: rewrite stale diffusion compatibility matrix (#36987) 2026-09-02 23:42:54 +08:00
4b329482e8 [diffusion] feat: support cube sparse attention for minimax h3 (#34893)
Co-authored-by: zhenaozhenfu <zhenaozhenfu@minimaxi.com>
Co-authored-by: Reynor <reynor@minimaxi.com>
2026-09-02 15:21:33 +08:00
Mick 9175590aa0 [diffusion] refactor: admit explicit attention backends by capability (#37441) 2026-09-02 15:17:57 +08:00
Mick 1109e44305 update CODEOWNERS (#37529) 2026-09-02 12:19:11 +08:00
Mick 22195709d1 [diffusion] refactor: remove component loader capability switches (#36824) 2026-09-02 09:48:50 +08:00
Mick dde0ecdb90 [diffusion] feat: support spargeattention (#37437) 2026-09-02 09:42:33 +08:00
Mick ae2bd5728b [vlm] fix: contain multimodal feature transport failures (#37047) 2026-09-01 13:46:38 +08:00
Mick 2cb3f32b03 [diffusion] chore: preserve exact component identity during loading (#36875) 2026-08-31 14:17:57 +08:00
Mick 62c470697e [diffusion] chore: enforce component attention backend application (#36907) 2026-08-31 14:00:02 +08:00
Mick 28690f5aa5 [diffusion] chore: detect quantized transformer replacements (#36916) 2026-08-31 13:55:50 +08:00
Mick 881cbfe54c [diffusion] feat: add exact component precision overrides (#36991) 2026-08-31 11:12:07 +08:00
Mick fe694986a2 [diffusion] chore: make malformed component execution options fail-fast (#37049) 2026-08-30 21:06:19 +08:00
Mick e6a6492057 [vlm] fix: preserve per-request vit graph metadata for qwen-vl (#37043) 2026-08-30 21:02:14 +08:00
Mick 26c754e06e [vlm] fix: recover multimodal decode and processor failures (#36983) 2026-08-30 20:50:19 +08:00
Mick aa483ab782 [diffusion] feat: support streaming native vae weights directly to gpu (#37004) 2026-08-30 20:48:09 +08:00
Mick a6e4021368 [diffusion] chore: reject incompatible transformer fallback (#36917) 2026-08-30 16:33:14 +08:00
Mick 97781eb7f3 [diffusion] fix: honor explicit offload in resident requirements (#36905) 2026-08-29 22:48:31 +08:00
Mick b24bd44556 [diffusion] feat: avoid direct GPU parameter copies (#36832) 2026-08-29 22:43:25 +08:00
Mick fa474b0441 [diffusion] fix: fix image encoder parallel folding proposal (#36863) 2026-08-29 14:34:36 +08:00
Mick f5319af1c9 [diffusion] chore: honor explicit component offload (#36931) 2026-08-29 14:29:57 +08:00
Mick d1ce017665 [diffusion] feat: delegate recognized quantized components to transformers (#36902) 2026-08-29 14:26:50 +08:00
Mick 3c1d77be21 [diffusion] feat: respect component weight overrides for upsamplers (#36874) 2026-08-29 14:23:11 +08:00
Mick f93b48c627 [diffusion] refactor: resolve indexed component weight sets (#36883) 2026-08-29 14:21:25 +08:00
Mick 803b4fb31c [diffusion] refactor: scope model-specific API parameters (#35613) 2026-08-28 19:08:30 +08:00
Mickandjohn.xh 20a491d1d3 [Fix] Re-encode multimodal embeddings after cache mismatch (#36595)
Co-authored-by: john.xh <dengxuhuijohn@gmail.com>
2026-08-27 23:24:30 +08:00
MickandClaude Opus 5 c2c3320cf0 [VLM] feat: size the multimodal preprocessing pool by where preprocessing runs (#35349)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-08-27 16:32:58 +08:00
MickandClaude Opus 5 07570202f0 [VLM] route every multimodal processor through the worker pool's call site (#35342)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-08-27 10:27:45 +08:00
Mick c3c529c28e [diffusion] docs: distinguish MiniMax H3 checkpoint variants (#36412) 2026-08-26 15:16:18 +08:00
Mick 054f485d38 [diffusion] docs: add MiniMax H3 checkpoint format table (#36028) 2026-08-26 09:18:07 +08:00
Mick 223dfce917 fix: bound CUDA memory for fast image preprocessing (#36295) 2026-08-26 09:02:08 +08:00
MickandClaude Fable 5 c3947eeada [diffusion] docs: desktop-safe 24 GB recipe and the DGX Spark tier (#36169)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-25 20:25:17 +08:00
Mick 191244b3f6 [diffusion] feat: support loading Comfy NVFP4-AWQ text encoders (#36046) 2026-08-25 15:43:30 +08:00
Mick 67853c5804 [diffusion] feat: dispatch fp8 companions in mixed NVFP4 checkpoints (#36066) 2026-08-25 11:26:37 +08:00
Mick f8f9226cd2 [diffusion] feat: support component-scoped quantization overrides (#36035) 2026-08-25 09:20:47 +08:00
Mick ddea7b9156 [diffusion] feat: support mixed Comfy NVFP4 and INT8 layers (#36061) 2026-08-25 09:19:47 +08:00
Mick 30f9ed09d1 [diffusion] feat: support loading minimax h3 gguf text encoders (#36055) 2026-08-24 23:00:49 +08:00
Mick 9b0007ed19 [diffusion] feat: support loading comfy nvfp4 minimax h3 checkpoints (#36044) 2026-08-24 22:57:02 +08:00
Mick 76d1401881 [diffusion] feat: support mixed w4a4 and int8 checkpoints (#36040) 2026-08-24 20:35:11 +08:00
Mick 9856b58de4 [diffusion] feat: support loading serialized fp8 clip image encoders (#36056) 2026-08-24 19:59:35 +08:00
Mick bfeae4e79a [diffusion] feat: support loading serialized convrot w4a4 checkpoints (#36039) 2026-08-24 19:13:35 +08:00
Mick 51b27f747a [diffusion] feat: support loading pruned minimax h3 components natively (#36070) 2026-08-24 16:39:37 +08:00
Mick adc09a1f63 [diffusion] feat: support loading self-describing quanto int8 encoders (#36052) 2026-08-24 16:37:09 +08:00
Mick 5081ad5d4e [diffusion] feat: add per-component quantization overrides (#36084) 2026-08-24 16:35:48 +08:00
MickandClaude Fable 5 c8e1ddc707 [diffusion] feat: cache LoRA-merged weights in files the page cache can hold (#36062)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-24 14:18:55 +08:00
Mick 3fe18f13cd [diffusion] feat: add plain component weight overrides (#36086) 2026-08-24 13:33:23 +08:00
Mick 8df3b9eff9 [diffusion] feat: support loading mixed w4a8 text encoders (#36037) 2026-08-24 13:31:27 +08:00
Mick 7a7b655ddf [quantization] share bounded post-load device staging (#35180) 2026-08-24 11:48:30 +08:00
Mick 5ce700aee8 [diffusion] feat: infer LoRA alpha from safetensors metadata (#36082) 2026-08-24 11:46:18 +08:00
Mick 230c052ebc [diffusion] chore: reuse srt AutoRound for quantized DiTs (#36068) 2026-08-24 11:21:44 +08:00
MickandClaude Opus 5 f294d51a71 [diffusion] fix: fix a refit key error on mapped weights, and stop claiming strides the reload discards (#35832)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-08-24 11:11:35 +08:00
Mick 6ca872a11f [diffusion] chore: fetch metadata beside nested lora weights (#36057) 2026-08-24 11:02:23 +08:00
Mick 1a368eca1c [diffusion] optimization: reuse minimax h3 prompt refinement across outputs (#36027) 2026-08-24 11:01:21 +08:00
Mick f6fff25756 [diffusion] feat: support vae weight-file overrides (#36085) 2026-08-24 10:33:14 +08:00
Mick 2d84de5e69 [diffusion] feat: support loading serialized comfy w4a8 checkpoints (#36036) 2026-08-24 10:32:21 +08:00
Mick fee00a41db [diffusion] feat: add composable component weight path cli (#36078) 2026-08-24 09:31:08 +08:00
Mick 1c1c9d9b4e [diffusion] refactor: reuse srt quantization contracts and mxfp8 kernels (#36063) 2026-08-24 09:26:54 +08:00
Mick d1af3c8923 [diffusion] feat: support loading native diffusers miniMax h3 components (#36067) 2026-08-23 22:13:59 +08:00
Mick de6a1dbd7a [diffusion] feat: support hybrid conditioning for minimax h3 (#36080) 2026-08-23 21:51:25 +08:00
Mick 939c00a7e3 [diffusion] feat: support compact qwen3-vl conditioning for minimax h3 (#36076) 2026-08-23 21:25:11 +08:00
MickandClaude Fable 5 886e37a649 [diffusion] CI: guard the anonymous-host budget alongside peak VRAM (#36051)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-23 20:13:46 +08:00
Mick dd15fb57b5 [diffusion] feat: automatically infer comfy fp8 activation scaling (#36060) 2026-08-23 18:43:33 +08:00
Mick e3a008a9db [diffusion] UX: clean up startup and offload logs (#36034) 2026-08-23 16:50:25 +08:00
MickandClaude Fable 5 bd3cc97e7e [diffusion] CI: let the 5090 consumer case runs two warm requests on the full recipe (#36032)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-23 15:39:32 +08:00
Mick 97ae27893a [diffusion] feat: release a layerwise component's non-layer weights between uses (#35734) 2026-08-23 12:05:24 +08:00
Mick 70319a0881 [diffusion] feat: support loading serialized comfy convrot int8 native encoders (#36023) 2026-08-23 12:00:38 +08:00
Mick bbbcbf9418 [diffusion] fix: stabilize ltx-2.3 two-stage cold requests (#35997) 2026-08-23 10:04:00 +08:00