FanLR commited on
Commit
360e1c4
·
verified ·
1 Parent(s): d807bbf

Record two successful private ZeroGPU grounding checks

Browse files
Files changed (2) hide show
  1. README.md +5 -4
  2. docs/HF_PRIVATE_SMOKE.md +17 -0
README.md CHANGED
@@ -22,8 +22,8 @@ interaction targets. Choose a task, provide an image and a prompt, then inspect
22
  the model's predictions directly on the image.
23
 
24
  Model: [XPENG-AI/GroundingPI-4B](https://huggingface.co/XPENG-AI/GroundingPI-4B).
25
- This private preview is hosted at `FanLR/GroundingPI-4B`. HF ZeroGPU inference
26
- acceptance remains pending. Licensing and asset permissions are pending:
27
  see [license status](LICENSE_PENDING.md).
28
 
29
  ## Explore visual evidence
@@ -95,8 +95,9 @@ See [runtime design](docs/ZEROGPU_DEPLOYMENT.md).
95
 
96
  The existing PPU implementation has completed real generation and UI checks.
97
  This package includes automated regressions for prompts, parsing, stale-result
98
- clearing, error recovery and overlay geometry. **NVIDIA / HF ZeroGPU acceptance
99
- remains pending.** See [verification](docs/VERIFICATION.md).
 
100
 
101
  The default queue is single-concurrency with at most eight waiting requests.
102
  Uploads are limited to 30 MiB / 50 megapixels, and input plus output to 8,192 tokens.
 
22
  the model's predictions directly on the image.
23
 
24
  Model: [XPENG-AI/GroundingPI-4B](https://huggingface.co/XPENG-AI/GroundingPI-4B).
25
+ This private preview is hosted at `FanLR/GroundingPI-4B`. Two real HF ZeroGPU grounding requests passed on 2026-09-16; full task coverage
26
+ remains pending. Licensing and asset permissions are pending:
27
  see [license status](LICENSE_PENDING.md).
28
 
29
  ## Explore visual evidence
 
95
 
96
  The existing PPU implementation has completed real generation and UI checks.
97
  This package includes automated regressions for prompts, parsing, stale-result
98
+ clearing, error recovery and overlay geometry. **Two private HF ZeroGPU grounding smoke checks passed on 2026-09-16.**
99
+ See [the smoke-check receipt](docs/HF_PRIVATE_SMOKE.md). Full route coverage and
100
+ benchmark accuracy remain unverified. See [verification](docs/VERIFICATION.md).
101
 
102
  The default queue is single-concurrency with at most eight waiting requests.
103
  Uploads are limited to 30 MiB / 50 megapixels, and input plus output to 8,192 tokens.
docs/HF_PRIVATE_SMOKE.md ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Private HF ZeroGPU smoke checks
2
+
3
+ Verified on 2026-09-16 against model commit `9dd5c58abb69e87a9c8aa59865ed6bec944026ed`.
4
+ The model and Space remained private. Two sequential requests used the same
5
+ image from the third grounding example, with different user queries. Only
6
+ the image and query entered inference; no expected answer was supplied.
7
+
8
+ | Query | Generated box (0–999) | Demo-reported inference time |
9
+ | --- | --- | --- |
10
+ | Person | 463, 270, 888, 997 | 6.46 s |
11
+ | Keyboard | 362, 615, 571, 739 | 0.58 s |
12
+
13
+ Both requests reached EOS, passed native GAM parsing, and returned rendered
14
+ image overlays. The boxes were visually checked against the input image.
15
+ Reported times are individual observations and exclude the local client network
16
+ overhead; they are not a throughput benchmark. These checks cover basic
17
+ grounding only. Other routes and benchmark accuracy remain unverified.