Hugging Face
Search models, datasets, users...
Ctrl K
Models
Datasets
Spaces
Buckets
Docs
Enterprise
Pricing
Website
Tasks
HuggingChat
Collections
Languages
Organizations
Community
Blog
Posts
Daily Papers
Hardware
Learn
Discord
Forum
GitHub
Solutions
Team & Enterprise
Hugging Face PRO
Enterprise Support
Inference Providers
Inference Endpoints
Storage Buckets
Log In
Sign Up
Spaces:
cazyundee
/
training
Like
0
Running
on
Zero
App
Files
Files
Community
Fetching metadata from the HF Docker repository...
main
training
/
experiments
/
gemma_decoder
61.6 kB
Ctrl+K
Ctrl+K
2 contributors
History:
19 commits
agent
targets: two new grounded claims from the trace audit (idle memory; same-squares-keep-changing) - for the next pilot round
ca231d1
2 days ago
REPORT_DRAFT_notes.md
Safe
2.78 kB
prototype report skeleton with all verified facts (results section awaits Space pilot)
2 days ago
bridge.py
Safe
8.62 kB
NaN fix v2 (root cause: exactly-zero embedding rows trigger gemma-backward NaN on torch 2.13, localised to layer-12 mlp grad_out via hooks; A-probe with zero token reproduced it): no zero rows anywhere - B uses 1e-4 no-info epsilon, resampler opener 1e-4
2 days ago
capture.py
Safe
11.1 kB
gemma_decoder prototype: faithful live capture (hooks: router, mem_attn; ACT semantics), Latent+Resampler bridges (Gemma embed-scale aware, one-way detach boundary), grounded NL targets, run_pilot with controls A-G and frozen-Gemma gradient contract; 8/8 tests
2 days ago
debug_backend.py
Safe
2.59 kB
debug_backend: len-mismatch fix (eos slicing)
2 days ago
debug_d.py
Safe
2.59 kB
debug_d: restore m,cfg assignment lost in the hub-pull patch
2 days ago
debug_nan.py
Safe
4.81 kB
debug_nan: hook-based NaN localisation (bridge + all gemma layers) + zero-token and sdpa-shape probes
2 days ago
run_pilot.py
Safe
15.4 kB
NaN fix v2 (root cause: exactly-zero embedding rows trigger gemma-backward NaN on torch 2.13, localised to layer-12 mlp grad_out via hooks; A-probe with zero token reproduced it): no zero rows anywhere - B uses 1e-4 no-info epsilon, resampler opener 1e-4
2 days ago
run_pilot_v6.py
Safe
1.48 kB
run_pilot_v6: 96 pos x 150 iters, PT+IT, results auto-pushed to hub (rebuild-proof)
2 days ago
targets.py
Safe
7.04 kB
targets: two new grounded claims from the trace audit (idle memory; same-squares-keep-changing) - for the next pilot round
2 days ago
tests_gemma.py
Safe
5.12 kB
gemma_decoder prototype: faithful live capture (hooks: router, mem_attn; ACT semantics), Latent+Resampler bridges (Gemma embed-scale aware, one-way detach boundary), grounded NL targets, run_pilot with controls A-G and frozen-Gemma gradient contract; 8/8 tests
2 days ago