* huggingface-spaces: add 3D generation model coverage
Extend the Spaces skill to cover building Spaces for 3D generation
models: image-to-3D / text-to-3D mesh generation (TRELLIS/TRELLIS.2,
Hunyuan3D-2/2.1, TripoSR, Stable Fast 3D, SPAR3D) and gaussian
splatting (TripoSplat, LGM, Apple SHARP, Splatter Image), including
finetuned checkpoints of those models.
New references:
- 3d-generation.md — workflow: model selection, license/gating checks,
duplicating official Spaces vs building from source, 3D verification
- 3d-models.md — per-family recipes (TRELLIS family; TripoSR, SF3D,
SPAR3D, Hunyuan3D) distilled from the live files of the official
Spaces as of 2026-07
- 3d-gsplat.md — gaussian splatting models, ply/splat formats,
rasterize-or-not decision
- 3d-cuda-extensions.md — the no-nvcc-at-build problem and the three
extension strategies (prebuilt wheels, committed wheels, JIT
bootstrap), durations, GPU-boundary rules for meshes/gaussians
- 3d-outputs.md — GLB/OBJ/PLY/splat formats, gr.Model3D / LitModel3D
viewers, orientation fixes, preprocessing patterns
The mesh path was validated end-to-end by duplicating
stabilityai/TripoSR to a private ZeroGPU Space and running a real
inference with trimesh mesh validation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* huggingface-spaces: headless Blender patterns and HF Jobs wheel building
Two additions to the 3D references, verified against running Spaces:
- 3d-outputs.md: new "Headless Blender" section — pip bpy and modern
(>=3.4) Blender binaries run display-free via EGL (packages.txt
recipes from working bpy Spaces); xvfb is the fallback for pinned
pre-3.4 binaries, EEVEE on legacy setups, or Open3D/VTK/pyglet GL
stacks, via pyvirtualdisplay on the gradio SDK (raw Xvfb hits
/tmp/.X11-unix permissions) or xvfb-run on Docker, as used by
gradient-spaces/GuideFlow3D to render Cycles through Blender 3.0.1;
ZeroGPU gotcha: import bpy takes 30-60s, keep it out of @spaces.GPU;
pyrender should use EGL, not xvfb.
- 3d-cuda-extensions.md: build missing extension wheels on HF Jobs
(pytorch devel image matching the ZeroGPU runtime, TORCH_CUDA_ARCH_LIST
for sm_120, hf upload to a Hub repo; the rtx-pro-6000 flavor is the
same Blackwell silicon so the job can smoke-test the kernel), then
reference the wheel by resolve URL or commit it to the Space.
Also corrects the Hunyuan3D-2.1 bpy note: bpy==4.0 installs cleanly
headless, but that Space's GLB conversion actually uses
trimesh+pygltflib.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: bump version num for metadata
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: bump metadata versions and sync
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: remove disproportionate 3d information
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: remove agents.md references for 3d
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: repair agents.md and skills.md char alignment
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: remove 3d mentions from agents and skills markdown
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: repair agents.md file
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: version bump
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: remove per-sdk depth info
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* Update skills/huggingface-spaces/SKILL.md
Co-authored-by: apolinário <joaopaulo.passos@gmail.com>
* chore: address comments
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* Update AGENTS.md
Co-authored-by: apolinário <joaopaulo.passos@gmail.com>
---------
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
Co-authored-by: Suvaditya Mukherjee <suvaditya@hf.co>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: apolinário <joaopaulo.passos@gmail.com>
* huggingface-spaces: add 3D generation model coverage
Extend the Spaces skill to cover building Spaces for 3D generation
models: image-to-3D / text-to-3D mesh generation (TRELLIS/TRELLIS.2,
Hunyuan3D-2/2.1, TripoSR, Stable Fast 3D, SPAR3D) and gaussian
splatting (TripoSplat, LGM, Apple SHARP, Splatter Image), including
finetuned checkpoints of those models.
New references:
- 3d-generation.md — workflow: model selection, license/gating checks,
duplicating official Spaces vs building from source, 3D verification
- 3d-models.md — per-family recipes (TRELLIS family; TripoSR, SF3D,
SPAR3D, Hunyuan3D) distilled from the live files of the official
Spaces as of 2026-07
- 3d-gsplat.md — gaussian splatting models, ply/splat formats,
rasterize-or-not decision
- 3d-cuda-extensions.md — the no-nvcc-at-build problem and the three
extension strategies (prebuilt wheels, committed wheels, JIT
bootstrap), durations, GPU-boundary rules for meshes/gaussians
- 3d-outputs.md — GLB/OBJ/PLY/splat formats, gr.Model3D / LitModel3D
viewers, orientation fixes, preprocessing patterns
The mesh path was validated end-to-end by duplicating
stabilityai/TripoSR to a private ZeroGPU Space and running a real
inference with trimesh mesh validation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* huggingface-spaces: headless Blender patterns and HF Jobs wheel building
Two additions to the 3D references, verified against running Spaces:
- 3d-outputs.md: new "Headless Blender" section — pip bpy and modern
(>=3.4) Blender binaries run display-free via EGL (packages.txt
recipes from working bpy Spaces); xvfb is the fallback for pinned
pre-3.4 binaries, EEVEE on legacy setups, or Open3D/VTK/pyglet GL
stacks, via pyvirtualdisplay on the gradio SDK (raw Xvfb hits
/tmp/.X11-unix permissions) or xvfb-run on Docker, as used by
gradient-spaces/GuideFlow3D to render Cycles through Blender 3.0.1;
ZeroGPU gotcha: import bpy takes 30-60s, keep it out of @spaces.GPU;
pyrender should use EGL, not xvfb.
- 3d-cuda-extensions.md: build missing extension wheels on HF Jobs
(pytorch devel image matching the ZeroGPU runtime, TORCH_CUDA_ARCH_LIST
for sm_120, hf upload to a Hub repo; the rtx-pro-6000 flavor is the
same Blackwell silicon so the job can smoke-test the kernel), then
reference the wheel by resolve URL or commit it to the Space.
Also corrects the Hunyuan3D-2.1 bpy note: bpy==4.0 installs cleanly
headless, but that Space's GLB conversion actually uses
trimesh+pygltflib.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: bump version num for metadata
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
* chore: bump metadata versions and sync
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
---------
Signed-off-by: Suvaditya Mukherjee <suvaditya@hf.co>
Co-authored-by: Suvaditya Mukherjee <suvaditya@hf.co>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Fix gr.Examples fn/inputs signature guidance in lora-space-builder
Add the rule that gr.Examples(fn=...) calls fn with only the inputs=
values positionally, so uncovered params need defaults or the first
example click raises a missing-positional-argument TypeError.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* bump!
* bump!
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: evalstate <1936278+evalstate@users.noreply.github.com>
* add: Hugging Face skills for the AWS cloud provider
Six hf-cloud-* skills for deploying models to Amazon SageMaker:
deployment planner, AWS context discovery, Python env setup, IAM
preflight, serving-image selection, and production endpoint defaults.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore: regenerate skill metadata for hf-cloud skills
Regenerates AGENTS.md, README table, and marketplace-internal.json to
include the six hf-cloud-* skills; bumps plugin version to 1.0.14.
Public marketplace.json stays curated to hf-cli.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* Update huggingface-spaces skill: wheels matrix, AoTI, attention, themes
- requirements.md: real 12-cell Blackwell wheels matrix; flash_attn cp310-cp313
(2.8.3 + 2.7.4.post1); per-cell xformers filename caveat; torch-minor tolerance
- zerogpu.md attention: SDPA first-class default, FA2 only when the repo uses FA;
document FA2 as the ceiling (FA3/FA4 can't run on sm_120 — no TMEM)
- zerogpu.md AoTI: rewrite around the current spaces>=0.50 API
(aoti_capture / aoti_compile_and_save / aoti_blocks_load) and the precompile-block
-> Hub -> load two-Space workflow, with footguns and reference Spaces
- SKILL.md: instruct looking up the current latest Gradio, not the placeholder
- gradio.md: prefer gr.themes.Citrus() + required .dark .gradio-container CSS fix
* huggingface-spaces: require --repo-type space on hf upload; exclude __pycache__
hf upload defaults to a model repo, so Space uploads must pass --repo-type
space or they upload to (and silently create) a model repo of the same name.
Fix the debugging.md rung-2/3 commands and the SKILL.md iterate section, and
add --exclude "**/__pycache__/**" so local bytecode caches aren't committed.
* huggingface-spaces: always add gr.Examples, docstrings, and MCP server
- gr.Examples: prescriptive "add whenever inputs exist", prefer the original
repo/model's official examples, and keep example rows minimal by giving the
handler default params (examples vary only the key inputs); cache_examples=True,
cache_mode="lazy"
- MCP: launch(mcp_server=True) to expose API endpoints as MCP tools; every
API-triggered function gets a docstring + type hints (become the tool schema).
New gradio.md subsection + SKILL.md §5 directives + minimal-app example updated
* Bump plugin version 1.0.10 -> 1.0.11
* merge and bump
---------
Co-authored-by: multimodalart <joaopaulo.passos+multimodal@gmail.com>
Co-authored-by: evalstate <evalstate@users.noreply.github.com>
* docs: document private= / bucket_id= in trackio examples; update stale Jobs sync description
Agents following the trackio examples create public Spaces by default
(examples show space_id= but never private=). Document private and
bucket_id in the huggingface-trackio skill and the llm-trainer trackio
guide, framing private=True as a choice (omit for a shareable public
dashboard).
Also update the llm-trainer guide's 'How It Works with Jobs' section,
which described the old Dataset/Parquet 5-minute sync: current trackio
streams to the running Space (batched sub-second), spills to an HF
Bucket every ~30s when the Space is unreachable/building, and the Space
ingests the bucket inbox every ~15s (verified against trackio main:
run.py BATCH_SEND_INTERVAL/BUCKET_FLUSH_INTERVAL, utils.get_inbox_poll_interval).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: bump plugin versions to 1.0.11
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* version bump
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: shaun smith <1936278+evalstate@users.noreply.github.com>
Co-authored-by: evalstate <evalstate@users.noreply.github.com>
* feat(distribution): emit SEP-2640 index format + full expanded tree
Update build_skill_distribution.py to the latest Skills Over MCP WG index
schema (decisions 2026-06-05 / 2026-06-09):
- index.json entries now carry verbatim `frontmatter`, `url` + `digest`
(sha256 of SKILL.md), and a per-skill `archives[]` array. Drop the old
`name`/`type`/`description`/`$schema` 0.2.0 discovery fields.
- Lay out the full expanded skill tree under <out>/<name>/... (every file,
not just SKILL.md) alongside the .tar.gz, so consumers can address each
file individually and walk directories (resources/directory/read).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* either SKILL.md OR .tar.gz.
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: evalstate <1936278+evalstate@users.noreply.github.com>
* fix: add required "type": "http" to .mcp.json
Claude Code validates MCP server configs against a discriminated union (stdio | sse | http). A bare {url} entry matches no branch and is rejected at load time with 'Invalid MCP server config: invalid_union', so the server never registers for plugin users. All documented HTTP examples carry an explicit "type": "http".
* update generator to include type
* Bump plugin version for MCP config change
* Rename generated agents directory to agentsmd
---------
Co-authored-by: whatrwewaitingf0r <948960638175457471+whatrwewaitingf0r@users.noreply.github.com>
Dario Salvati