Mar 28 10:18 rls t5 large 08:55 rls t5 large 05:09 rls 02:42 rls Mar 27 11:54 rls 09:35 new v3 Mar 24 16:52 v10a 16:48 v10a 16:44 v10a 16:23 v10a 16:17 v10a 16:03 v10a 15:48 v10a 08:35 V10a 07:49 V10 04:18 v10 04:02 v10 03:45 v10 03:25 v10 Mar 23 19:27 v10 18:10 v10 04:57 v9 04:55 v9-fix: oracle training routing, calibrated inference, update docs 04:28 v9 Mar 22 19:08 v8 05:13 v7 04:10 v7-fix: enable C5 for task 2+ (add --run_single False, --n_batches_c5 100) Mar 21 05:18 v7 05:17 v7 04:40 v7: C5 Data-Informed Subspace Init + restructure contributions to 2 core claims Mar 18 09:35 fix: reduce CPU RAM to prevent OOM SIGKILL - gc.collect, del temps, eval_accumulation_steps, cache cleanup 09:01 Revert "fix: reduce OOM in V2 script - fp16+gradient_ckpt for T4, EVAL_BSZ 128->32" This reverts commit 0289639e86234ac68d3468656223c6b3abc85044. 08:56 fix: reduce OOM in V2 script - fp16+gradient_ckpt for T4, EVAL_BSZ 128->32 - Enable --fp16 for T4 GPUs (was fp32 despite comment saying gradient_ckpt)
- Enable --gradient_checkpointing True for T4 GPUs
- Reduce EVAL_BSZ from 128 to 32 for T4 (2gpu and 1gpu configs)
- A100 config unchanged (EVAL_BSZ=128 fine with high VRAM)
Root causes of OOM at task 10 (dbpedia, 9 prev LoRA paths):
1. EVAL_BSZ=128 with seq_len=512: encoder attention maps (BSZ,8,512,512) spike VRAM
2. gradient_checkpointing was mentioned in comment but never actually enabled
3. fp32 on T4 when fp16 is fully supported (halves activation/model memory) 02:13 fix: add trust_remote_code=True to second load_dataset call in run_llama.py 01:08 dataset: allow custom dataset code by setting trust_remote_code=True in run_llama.py Mar 17 17:15 C4: Spectrally-Conditioned LoRA Training — preconditioned gradient + spectral entropy regularization - cl_trainer_specroute.py: precompute_preconditioners(), _compute_spectral_entropy_loss(), _apply_preconditioning()
- run_t5.py: 4 new args (lambda_entropy, use_preconditioning, precond_eps, entropy_warmup_ratio)
- V4 shell script: gen_script_long_order3_t5_small_specroute_v4.sh
- SPECROUTE_IDEA.md: C4 section, Claim 4, updated tables
- experiment_versions.md: V4.0 section + ablation plan
- contribution2_implementation_analysis.md: detailed C4 analysis (math, pipeline, synergy, risks) 16:43 SpecRoute V3: adaptive bias, symmetric inference, threshold 0.995, batch size optimization V3 fixes (3 targeted):
- Adaptive training bias: beta=T*ln(alpha*n/(1-alpha)), ensures ~80% routing weight
- Symmetric inference routing: prepare_inference_routing() computes SVD for current task
- Threshold 0.995: matches ROOT GPM capacity
Batch size optimization:
- T4_1gpu: BSZ 4->8, GA 8->4 (effective=32, ~2x fewer micro-batches)
- T4_2gpu: BSZ 2->4, GA 8->4 (effective=32)
- A100: BSZ 8->32, GA 4->1 (effective=32)
Docs: V2 diagnosis report, V3 methodology in SPECROUTE_IDEA.md, experiment_versions.md Mar 16 11:50 fix: eval_strategy epoch (was steps=10, causing 437 evals/task → Kaggle session kill) 10:53 fixbug 10:42 fixbug 10:08 fixbug 10:06 fixbug 09:08 fix llama 08:48 fix llama 
