Integrate mistral-common for correct Devstral tokenization
ed06dcb
gary-boon
Claude Opus 4.5
commited on
Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
Commit History
Use mistral_common for proper Devstral prompt formatting
3e80769
gary-boon
Claude Opus 4.5
commited on
Add system prompt support for instruction-tuned models
2860768
gary-boon
Claude Opus 4.5
commited on
fix: Simpler prompt format and temperature=0 for Devstral
76020ee
gary-boon
Claude Opus 4.5
commited on
fix: Sanitize JSON response for NaN/Inf float values
99f6209
gary-boon
Claude Opus 4.5
commited on
fix: Check chat_template is set before using apply_chat_template
474927d
gary-boon
Claude Opus 4.5
commited on
fix: Add chat template support for Devstral instruct model
8d85da8
gary-boon
Claude Opus 4.5
commited on
fix: Convert bfloat16 to float32 for numpy compatibility
cb6f39c
gary-boon
Claude Opus 4.5
commited on
fix: Use eager attention for output_attentions support
5333b21
gary-boon
Claude Opus 4.5
commited on
Add vocabSize to modelInfo response
499afba
gary-boon
Claude Opus 4.5
commited on
Add recommended_dtype to model configs
62525b2
gary-boon
Claude Opus 4.5
commited on
Phase 2: Add Devstral backend support
9080f28
gary-boon
Claude Opus 4.5
commited on
Add DEVICE env var to force CPU mode on DGX Spark
5f122aa
gary-boon
Claude Opus 4.5
commited on
Phase 1: DGX Spark infrastructure
a2bd186
gary-boon
Claude Opus 4.5
commited on
Make QKV hook robust against shape mismatches
343dd57
gary-boon
Claude
commited on
Fix research attention endpoint model compatibility
f5ba954
gary-boon
Claude
commited on
Add research attention analysis endpoint with real CodeGen tokenization
8f63685
gary-boon
Claude
commited on
Add research attention analysis endpoints with Q/K/V extraction
37ed739
gary-boon
Claude
commited on
Fix ablation study for Code Llama compatibility
cd300ee
gary-boon
Claude
commited on
Fix model info endpoint for Code Llama compatibility
7dd568f
gary-boon
Claude
commited on
Add Code Llama 7B support with hardware-aware filtering and ICL timeout fixes
ed40a9a
gary-boon
Claude
commited on
Fix pyarrow compatibility issue with datasets library
1680fda
gary-boon
Claude
commited on
Add SWE-bench integration and improve backend routing
4444ae2
gary-boon
Claude
commited on
Add layer_stride parameter for PromptDiff optimization
5aed1a9
gary-boon
Claude
commited on
Capture complete attention patterns after generation
992dc8c
gary-boon
commited on
Fix: Use scaling approach instead of skipping layers
3c774b5
gary-boon
Claude
commited on
Fix: Refine layer hook output format handling
4b03268
gary-boon
Claude
commited on
Fix: Handle single-element tuple outputs in layer hook
9e42df9
gary-boon
Claude
commited on
Fix: Correct layer hook output format for layer_norm compatibility
070f9b8
gary-boon
Claude
commited on
Fix: Prevent hook persistence after ablation errors
3ee2b4b
gary-boon
Claude
commited on
feat: Add pipeline analyzer and QKV extractor for transformer visualization
767a3fd
gary-boon
Claude
commited on
Add backend support for ICL emergence analysis
920a98d
gary-boon
Claude
commited on
Add ablation support to model service with comprehensive testing
bb8a292
gary-boon
Claude
commited on
Add CORS support for Vercel production domain
3c5bd74
gary-boon
commited on
Add API key authentication
96a6300
gary-boon
commited on
Fix backend structure - remove duplicates
2c0fd9b
gary-boon
commited on