vrr/unsloth
2
0
Fork 0
mirror of https://github.com/unslothai/unsloth.git synced 2026-10-21 22:12:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 46

Commit graph

  • 7c63d8a811 Give the startup transfer budget headroom again after #14524 set it to its own build fix-bundle-budget-headroom Daniel Han 2026-10-21 15:09:53 -07:00
  • 3ee312cbbd Vary runs the selected image's recipe instead of the current form studio-images-sdxl-controls Daniel Han 2026-10-21 14:58:59 -07:00
  • 792723fe4d Studio: a multi-device sidecar launch keeps the largest card's share of the reserve studio-dspark-sidecar-fit-metal-oom Daniel Han 2026-10-21 15:03:43 -07:00
  • 5ab80ffdb4
         Studio: pin the master-982 stable-diffusion.cpp prebuilt so MOSS-Image GGUFs load (#14565) main Daniel Han 2026-10-21 15:02:19 -07:00
  • dcb5a8ed32 Warn when a literal-prompt image is restored onto the native sd.cpp engine Daniel Han 2026-10-21 14:58:58 -07:00
  • 24d2b9e5f7 Name imported variation seeds and strengths outside the request bounds as not applied Daniel Han 2026-10-21 14:58:28 -07:00
  • 8b75626079 Studio: panel residency uses the launch's whole-load proof everywhere; Metal residency prices the drafter; wording follows residency studio-lazy-table-fit Daniel Han 2026-10-21 14:58:11 -07:00
  • e906a12476 Build the CFG replica with captures held off, and fall back when the result copy home fails studio-cfg-parallel-default-tier Daniel Han 2026-10-21 14:58:04 -07:00
  • 435ac6907e Move Images-page-only gallery and ControlNet calls out of the eager api chunk Daniel Han 2026-10-21 14:53:16 -07:00
  • d0a65a69b8 Studio: unmasked spawns count every detected card, draft micro-batch capped at the batch, discrete Auto keeps a remote drafter Daniel Han 2026-10-21 14:48:26 -07:00
  • 2b7ebd3099 Copy CFG replica inputs with captures held off, and fall back to the source card when the copy fails Daniel Han 2026-10-21 14:46:58 -07:00
  • ea1168e5a4 Gate the CFG batch timing assertion on the latents' device, add a CUDA variant studio-diffusion-defaults-on Daniel Han 2026-10-21 14:38:59 -07:00
  • 34d54b040c Trim diffusion comments Daniel Han 2026-10-21 14:28:43 -07:00
  • 1e6be05dbf Merge remote-tracking branch 'origin/main' into r14562 Daniel Han 2026-10-21 14:40:45 -07:00
  • 31d484c0cc Hold captures off while a CFG replica result is copied to the primary card Daniel Han 2026-10-21 14:35:15 -07:00
  • 41ba5f13ea Release the capture hold when the CFG replica worker fails to start Daniel Han 2026-10-21 14:35:15 -07:00
  • a7fda46d8f Install the CUDA graph beneath a CFG router nested in a static step skip Daniel Han 2026-10-21 14:35:15 -07:00
  • 5ab5419a24
         Pin Lemonade 2026.43.3 (#14567) Daniel Han 2026-10-21 14:33:40 -07:00
  • 38d21245d0
         Studio Images: per-generation Shift for flow-matching models and SDXL textual inversion embeddings (#14563) Daniel Han 2026-10-21 14:29:44 -07:00
  • 4a625f84aa Studio: one-card flat reserve, repeated-device draft pin, compute floor for partly placed drafts, fitter minimum context Daniel Han 2026-10-21 14:28:27 -07:00
  • 4c44ffc9a7 Keep the imported prompt weighting check within the startup bundle budget Daniel Han 2026-10-21 14:26:35 -07:00
  • 3f22b3d2ef
         Studio: show the AMD NPU by default again, with a clearer setup notice (#14524) Daniel Han 2026-10-21 14:26:07 -07:00
  • 0a1316badc [pre-commit.ci] auto fixes from pre-commit.com hooks pre-commit-ci[bot] 2026-10-21 21:25:37 +00:00
  • 6e6e1790cb Carry the Vary seed and strength into the Generate anyway retry Daniel Han 2026-10-21 14:22:25 -07:00
  • 73104fb0dd Name prompt weighting as not applied when a weighted recipe is imported on a model without it Daniel Han 2026-10-21 14:21:59 -07:00
  • 4028232c23 Refuse prompt weighting off on the native sd.cpp engine, which always parses attention syntax Daniel Han 2026-10-21 14:21:35 -07:00
  • a0da256320 Merge remote-tracking branch 'origin/main' into studio-diffusion-defaults-on Daniel Han 2026-10-21 14:13:30 -07:00
  • c66b97d586 Trim comments in CFG-parallel graph changes Daniel Han 2026-10-21 14:11:28 -07:00
  • c27c9c61aa Studio: one resident list for the auto tensor pair; priced one-card probe; cancelled auto exits report cancelled studio-multi-gpu-tp Daniel Han 2026-10-21 14:09:37 -07:00
  • 94135efcad
         Take the least of three RssAnon peaks in the block dequant host RAM check test (#14561) Daniel Han 2026-10-21 14:11:04 -07:00
  • 5d60995af4 Studio diffusion: small DiTs kept bf16 compile from render 1 Daniel Han 2026-10-21 14:11:03 -07:00
  • a0b477368d Merge origin/main into studio-cfg-parallel-default-tier Daniel Han 2026-10-21 14:08:57 -07:00
  • c657d4bf91 Note that SDXL CLIP skip moves both text encoders, as in ComfyUI Daniel Han 2026-10-21 14:01:02 -07:00
  • 2d5c846481 Send variation seed fields only from Create Daniel Han 2026-10-21 14:00:32 -07:00
  • ddff6a3ec8 Name unsupported SD controls once in the not-applied toast Daniel Han 2026-10-21 14:00:13 -07:00
  • 08c4e21648 Restore Prompt weighting off from a gallery image made without it Daniel Han 2026-10-21 14:00:03 -07:00
  • 5549495814 Cap weighted prompt encoding at 32 chunks Daniel Han 2026-10-21 13:58:15 -07:00
  • e4aae5d233 Keep the opening parenthesis of a malformed prompt weight Daniel Han 2026-10-21 13:58:14 -07:00
  • 6e63ceb60d
         Studio: Z-Image CFG split under static step skip and CFG-parallel (#14551) Daniel Han 2026-10-21 14:02:35 -07:00
  • fa9082fe9d Studio: resident lookup tables follow the launch's fit on every retry and in the panel; probes leave host inputs out of VRAM Daniel Han 2026-10-21 13:58:23 -07:00
  • 81d812ff55 Studio tests: register MOSS-Image in the family-wide guards studio-moss-image Daniel Han 2026-10-21 13:51:42 -07:00
  • 76d91cfd21
         MoE autotune: run the kernel tuning eagerly when called from a compiled block (#14559) Daniel Han 2026-10-21 13:53:37 -07:00
  • 949a3c8087 Studio: refuse native-only MOSS-Image picks before the diffusers class check Daniel Han 2026-10-21 13:51:42 -07:00
  • 6db403a8ed [pre-commit.ci] auto fixes from pre-commit.com hooks pre-commit-ci[bot] 2026-10-21 20:47:07 +00:00
  • fa8cd6c363 Trim comments in SD controls Daniel Han 2026-10-21 13:41:51 -07:00
  • 35b4cf4ddd
         Count the context ceiling's per-device reserve gate from #14492 in the compute buffer wiring test (#14560) Daniel Han 2026-10-21 13:39:15 -07:00
  • 61e59cfc89 Studio: the residency estimate reads fp8 single files from the header too studio-diffusion-fp8-weight-storage Daniel Han 2026-10-21 13:30:03 -07:00
  • f5effabb00 Studio: hold graph recordings off for all CFG replica thread CUDA work Daniel Han 2026-10-21 13:29:14 -07:00
  • 870e58c61c Merge origin/main (#14508 paste parameters) into studio-images-sdxl-controls Daniel Han 2026-10-21 13:16:51 -07:00
  • 867f6306d0 Trim comments in MOSS-Image callback and preview tests Daniel Han 2026-10-21 13:11:44 -07:00
  • ba0ea9a8dc Studio: auto tensor pair stays off while a transformers model is resident; fit probe ranks discrete GPUs first Daniel Han 2026-10-21 13:02:10 -07:00
  • 94a8bfacb3
         Studio Images: paste generation parameters, inpaint mask blur and only-masked area, seed buttons (#14508) Daniel Han 2026-10-21 12:59:25 -07:00
  • 39ccf6bb74 Studio: live preview reads MOSS-Image [B, 128, h, w] latents with the FLUX.2 map Daniel Han 2026-10-21 12:56:08 -07:00
  • 2e2c0ffb8d Studio: gate LoRA off for MOSS-Image, whose pipeline has no LoRA loader Daniel Han 2026-10-21 12:56:08 -07:00
  • 2128da8870 Studio: build MOSS-Image step-end callback kwargs in the call scope Daniel Han 2026-10-21 12:56:08 -07:00
  • fe9aecf744 Studio: sidecar reserve follows the spawned child's devices, context and the user's margin Daniel Han 2026-10-21 12:53:54 -07:00
  • c5fc4a0004 Studio: keep CFG-parallel armed on the default speed tier and pair Z-Image branches Daniel Han 2026-10-21 12:45:01 -07:00
  • 1c0b578b4e [pre-commit.ci] auto fixes from pre-commit.com hooks pre-commit-ci[bot] 2026-10-21 19:27:43 +00:00
  • 02536afcf7 Trim comments in MOSS-Image support Daniel Han 2026-10-21 12:23:44 -07:00
  • 4a5a3e5828
         MoE grouped GEMM: per-expert dW kernel (#14550) Daniel Han 2026-10-21 12:11:05 -07:00
  • 80ef810a18 Merge remote-tracking branch 'origin/main' into studio-moss-image Daniel Han 2026-10-21 12:04:26 -07:00
  • 04f66ec703 Merge branch 'main' into trim-comments-backend-tests trim-comments-backend-tests Daniel Han 2026-10-21 11:57:49 -07:00
  • 7fe29903d7 Studio: auto tensor pair stays off while a GPU dictation sidecar is resident Daniel Han 2026-10-21 12:01:38 -07:00
  • db612da98e Studio diffusion: detect fp8 single files from the header, not the basename Daniel Han 2026-10-21 11:58:57 -07:00
  • a5f1cfeea1 Merge branch 'main' into trim-comments-studio-backend trim-comments-studio-backend Daniel Han 2026-10-21 11:49:16 -07:00
  • e4d57ae3c0 Studio: MOSS-Image keeps the bf16 denoiser by default while it fits Daniel Han 2026-10-21 11:57:23 -07:00
  • d617f7518c Studio: route MOSS-Image GGUFs to stable-diffusion.cpp Daniel Han 2026-10-21 11:50:43 -07:00
  • c105ac99a2 Merge branch 'main' into trim-comments-frontend-tests trim-comments-frontend-tests Daniel Han 2026-10-21 11:48:38 -07:00
  • bce4676519 Merge branch 'main' into trim-comments-backend-tests Daniel Han 2026-10-21 11:45:10 -07:00
  • ac267fbfc1 Merge branch 'main' into trim-comments-studio-backend Daniel Han 2026-10-21 11:29:43 -07:00
  • f42a3857ef Trim comments in fp8 storage code danielhanchen 2026-10-21 11:43:41 -07:00
  • d3969f71ae Studio diffusion: auto attention keeps native SDPA on Blackwell when it already runs cuDNN Daniel Han 2026-10-21 11:41:21 -07:00
  • 2aa50dbccb Studio diffusion: CFG batching for resident denoisers too, priced once per shape Daniel Han 2026-10-21 11:41:21 -07:00
  • a19151d03f Merge branch 'main' into trim-comments-unsloth-core trim-comments-unsloth-core Daniel Han 2026-10-21 11:16:38 -07:00
  • d22feecf22
         Studio: discrete-first GGUF placement keeps explicit device picks, matching context ceiling; in-process Rosetta check (#14492) Daniel Han 2026-10-21 11:25:24 -07:00
  • 5eaf2d84e1 Merge main into the lazy lookup-table branch Daniel Han 2026-10-21 11:23:36 -07:00
  • 3f74ad129f
         Studio: FLUX.3 upsampler pin follows torch's vendor, single-GPU hosts launch as before, non-blocking Vulkan probe (#14491) Daniel Han 2026-10-21 11:24:04 -07:00
  • f95c3a42f1 Studio: auto tensor pair fails closed on another resident's GPUs Daniel Han 2026-10-21 11:17:33 -07:00
  • 95dd6d6467 Studio: tighten draft pin match, read only the linked common library, keep the flat reserve for unsized sidecars Daniel Han 2026-10-21 11:20:32 -07:00
  • 9e6bec653c Merge origin/main: price fp8 storage in the single-file host RAM check Daniel Han 2026-10-21 11:17:15 -07:00
  • bf463ae39b Studio diffusion: store fp8 only where widening would offload or crowd host RAM Daniel Han 2026-10-21 11:13:55 -07:00
  • 3e92a375d3 Studio: the estimate reads a table resident where the launch would keep it in RAM Daniel Han 2026-10-21 11:12:48 -07:00
  • 33d163c314 Studio: price a DFlash/DSpark drafter's prompt graph with the micro-batch Daniel Han 2026-10-21 11:04:52 -07:00
  • ce1ab5b6a4
         Load the real loader_utils helpers into the precision conflict test harness (#14545) Daniel Han 2026-10-21 11:01:56 -07:00
  • b54cc6ac77 Studio: seed MOSS-Image's measured activation peak for low-VRAM placement Daniel Han 2026-10-21 10:56:02 -07:00
  • a7a37d3db2
         XPU bf16 on the current device, keep explicit bf16 on XPU, MLX full finetuning keeps its default dtype (#14493) Daniel Han 2026-10-21 10:55:35 -07:00
  • 56dcfe4859 [pre-commit.ci] auto fixes from pre-commit.com hooks pre-commit-ci[bot] 2026-10-21 17:53:21 +00:00
  • 1032b5f13b
         Studio: refuse a single-file diffusion load that would exhaust host RAM instead of crashing (#14510) Daniel Han 2026-10-21 10:52:58 -07:00
  • 427a59ade1 [pre-commit.ci] auto fixes from pre-commit.com hooks pre-commit-ci[bot] 2026-10-21 17:51:36 +00:00
  • 3f23fbbc7f [pre-commit.ci] auto fixes from pre-commit.com hooks pre-commit-ci[bot] 2026-10-21 17:49:47 +00:00
  • 0129f58c06 Studio: auto split retry honours an unload, Manual residents, flash-attn off Daniel Han 2026-10-21 10:49:25 -07:00
  • 7a9b36ee02 Studio tests: no placement stand-in server can reach a real process Daniel Han 2026-10-21 10:49:24 -07:00
  • b61583eed6 Merge branch 'main' into trim-comments-studio-frontend trim-comments-studio-frontend Daniel Han 2026-10-21 10:49:18 -07:00
  • d8d88aff9a
         Format the grouped GEMM single-flight wrapper so pre-commit.ci passes again (#14544) Daniel Han 2026-10-21 10:44:20 -07:00
  • b625a55423 Merge branch 'main' into trim-comments-installers-scripts trim-comments-installers-scripts Daniel Han 2026-10-21 10:43:13 -07:00
  • 11fac52f0d Merge branch 'main' into trim-comments-frontend-tests Daniel Han 2026-10-21 10:42:48 -07:00
  • 8aff9d7687 Merge branch 'main' into trim-comments-backend-tests Daniel Han 2026-10-21 10:38:23 -07:00
  • 5bca5ebe06 Merge branch 'main' into trim-comments-studio-backend Daniel Han 2026-10-21 10:26:18 -07:00
  • 11bedca900 [pre-commit.ci] auto fixes from pre-commit.com hooks pre-commit-ci[bot] 2026-10-21 17:36:59 +00:00
  • a0425b4dc6
         Keep two Backend CI tests valid on uv 0.13 and on runners with no accelerator (#14535) Daniel Han 2026-10-21 10:36:23 -07:00