Spaces:

Bhuvandesai
/

phi3-text-to-sql-studio

Sleeping

App Files Files Community

phi3-text-to-sql-studio

Commit History

Jul 9

15:48
fix: set per_device_eval_batch_size=1 to prevent OOM during evaluation loop

Bhuvandesai

26efb80
14:28
perf: increase batch size to 2 and enable group_by_length to massively speed up training

Bhuvandesai

c38b6f2
14:22
fix: use accelerate launch for DDP training instead of naive multi-GPU to avoid DataParallel crashes

Bhuvandesai

9638be2
14:19
fix: change device_map='auto' to local_process_index to prevent cross-GPU loss calculation error in transformers 4.46

Bhuvandesai

b602c40
14:16
fix: set remove_unused_columns=True so Trainer drops 'messages' list before data collation

Bhuvandesai

1bbece6
13:59
fix: change max_length to max_seq_length in SFTConfig for trl 0.12.1 compatibility

Bhuvandesai

930fd33
13:49
fix: restore PyTorch compatibility by removing triton downgrade, rely on latest bitsandbytes

Bhuvandesai

0139d8c
13:31
fix: downgrade triton to 3.1.0 in Kaggle notebook to fix bitsandbytes CUDA errors

Bhuvandesai

1735ce5
13:22
fix: upgrade bitsandbytes to >=0.45.0 to resolve Triton ModuleNotFoundError on Kaggle

Bhuvandesai

1cbd61f
12:17
feat: add Kaggle notebook for Phase 3 training

Bhuvandesai

021f693
12:13
docs(plan): deterministic rewrite — pin Spider to minktn/spider-data + anti-spiral rules; Phase 2 max_length=2560 + length filter⁠

Bhuvandesai +1: Claude Opus 4.8

d1cd1fa
12:11
refactor: address Phase 2 review findings⁠

Bhuvandesai +1: Claude Opus 4.8

e03c947
12:02
fix: LoRA adapts all 7 projections via Phi-3 fused names + guard assertion; r=16, Spider data, max_length=2560

Bhuvandesai

5f0c197
11:46
feat: Spider data pipeline pinned to minktn/spider-data (per-example CREATE TABLE schema); drop find_spider.py

Bhuvandesai

2ff9360
10:49
docs(plan): drop Phase 0 (repo reconciliation done via merge); point rollback at backup/pre-merge-1d8be28⁠

Bhuvandesai +1: Claude Opus 4.8

1bd448a
10:42
Merge remote-tracking branch 'origin/main'⁠

Bhuvandesai

25fa767
10:41
docs: condense writeup + add v2 implementation plan (Spider + fused-layer fix + CPU serving)⁠

Bhuvandesai +1: Claude Opus 4.8

670dcec

Jun 22

17:42
Complete CPU GGUF serving + docs + minimal UI redesign⁠

Bhuvandesai +1: Claude Opus 4.8

1d8be28
17:34
Redesign UI: flat minimal layout, no gradients/glass; add project stats panel and schema row counts

Bhuvandesai

Verified signature
61c1901
11:05
Rewrite CPU banner (T4 now paid; 5-8min load, 2-4min query) and add on-page FAQ

Bhuvandesai

Verified signature
c935f9c
10:54
Update Space README: GGUF/llama.cpp CPU serving, highlights, accurate timing

Bhuvandesai

Verified signature
9748758
10:49
Remove unreliable base-model comparison card; single fine-tuned result + wall-clock timer + accurate CPU banner

Bhuvandesai

Verified signature
9497393
10:33
Fix CPU thread oversubscription: cap n_threads to 2 (cpu-basic vCPUs) for faster generation

Bhuvandesai

Verified signature
4853f12
10:18
Fix build OOM (exit 137): CMAKE_BUILD_PARALLEL_LEVEL=1 (single-threaded compile)

Bhuvandesai

Verified signature
bf1a6c4
09:56
Fix inference SIGSEGV: pin llama-cpp-python==0.3.19 (0.3.2 crashed on this GGUF); verified locally

Bhuvandesai

Verified signature
2ee2903
09:36
Build llama-cpp-python from source (glibc); prebuilt wheels are musl-only for phi3-capable versions

Bhuvandesai

Verified signature
f438c10
09:26
Pin llama-cpp-python==0.3.19 + --only-binary (use prebuilt CPU wheel, no source build)

Bhuvandesai

Verified signature
2641362
08:57
Re-deploy llama.cpp + GGUF CPU serving (default Q4_K_M); fast CPU inference

Bhuvandesai

Verified signature
e44cdab

Jun 15

10:05
Revert "Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)"⁠

Bhuvandesai

d031aeb
10:05
Revert "Make torch/transformers/peft lazy imports so CPU Space boots without them"⁠

Bhuvandesai

2f03b0e
10:00
Make torch/transformers/peft lazy imports so CPU Space boots without them

Bhuvandesai

eac26c7
09:49
Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)

Bhuvandesai

564ad28
07:31
Fix bugs

Bhuvandesai

a7bb6b3
07:21
Fix bugs

Bhuvandesai

52e3716
06:57
Fix bugs

Bhuvandesai

19cfcb8
06:44
Update CPU warning banner with accurate load and query times

Bhuvandesai

e4ff54d
06:35
Fix inference device bug, auto-load model, remove fine-tuning console, UI cleanup

Bhuvandesai

f5ce94d
06:15
Add README with HF Spaces config

Bhuvandesai

6227517
06:12
initial deployment

Bhuvandesai

55159b1

Commit History

Jul 9

fix: set per_device_eval_batch_size=1 to prevent OOM during evaluation loop

perf: increase batch size to 2 and enable group_by_length to massively speed up training

fix: use accelerate launch for DDP training instead of naive multi-GPU to avoid DataParallel crashes

fix: change device_map='auto' to local_process_index to prevent cross-GPU loss calculation error in transformers 4.46

fix: set remove_unused_columns=True so Trainer drops 'messages' list before data collation

fix: change max_length to max_seq_length in SFTConfig for trl 0.12.1 compatibility

fix: restore PyTorch compatibility by removing triton downgrade, rely on latest bitsandbytes

fix: downgrade triton to 3.1.0 in Kaggle notebook to fix bitsandbytes CUDA errors

fix: upgrade bitsandbytes to >=0.45.0 to resolve Triton ModuleNotFoundError on Kaggle

feat: add Kaggle notebook for Phase 3 training

docs(plan): deterministic rewrite — pin Spider to minktn/spider-data + anti-spiral rules; Phase 2 max_length=2560 + length filter⁠

refactor: address Phase 2 review findings⁠

fix: LoRA adapts all 7 projections via Phi-3 fused names + guard assertion; r=16, Spider data, max_length=2560

feat: Spider data pipeline pinned to minktn/spider-data (per-example CREATE TABLE schema); drop find_spider.py

docs(plan): drop Phase 0 (repo reconciliation done via merge); point rollback at backup/pre-merge-1d8be28⁠

Merge remote-tracking branch 'origin/main'⁠

docs: condense writeup + add v2 implementation plan (Spider + fused-layer fix + CPU serving)⁠

Jun 22

Complete CPU GGUF serving + docs + minimal UI redesign⁠

Redesign UI: flat minimal layout, no gradients/glass; add project stats panel and schema row counts

Rewrite CPU banner (T4 now paid; 5-8min load, 2-4min query) and add on-page FAQ

Update Space README: GGUF/llama.cpp CPU serving, highlights, accurate timing

Remove unreliable base-model comparison card; single fine-tuned result + wall-clock timer + accurate CPU banner

Fix CPU thread oversubscription: cap n_threads to 2 (cpu-basic vCPUs) for faster generation

Fix build OOM (exit 137): CMAKE_BUILD_PARALLEL_LEVEL=1 (single-threaded compile)

Fix inference SIGSEGV: pin llama-cpp-python==0.3.19 (0.3.2 crashed on this GGUF); verified locally

Build llama-cpp-python from source (glibc); prebuilt wheels are musl-only for phi3-capable versions

Pin llama-cpp-python==0.3.19 + --only-binary (use prebuilt CPU wheel, no source build)

Re-deploy llama.cpp + GGUF CPU serving (default Q4_K_M); fast CPU inference

Jun 15

Revert "Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)"⁠

Revert "Make torch/transformers/peft lazy imports so CPU Space boots without them"⁠

Make torch/transformers/peft lazy imports so CPU Space boots without them

Migrate CPU serving to llama.cpp + Q5_K_M GGUF (was bf16 transformers)

Fix bugs

Fix bugs

Fix bugs

Update CPU warning banner with accurate load and query times

Fix inference device bug, auto-load model, remove fine-tuning console, UI cleanup

Add README with HF Spaces config

initial deployment