Resumable Kaggle evaluations, sequential SliceGPT LoRA and exported Modal progress

README.md

@@ -46,10 +46,13 @@ changes can still alter floating-point results.
 Existing Modal results/partial responses are available in modal_progress.zip.
 IMPORT_MODAL_PROGRESS=True imports them into the original BF16/32 profile.
 They are not reused for FP16/1 evaluations. Saved original pruning models and
-unfinished LoRA recovery files remain preserved. No paid Modal work runs.
+unfinished LoRA recovery files remain preserved. This portable runner never starts
+paid Modal compute.
 
 Original SliceGPT uploads are pruned baselines. None completed the requested
-sequential LoRA before the budget stop. Select a final fine-tuned repository
+sequential LoRA before the earlier evaluation budget stop. A separate capped
+fine-tuning campaign resumes those saved states; consult each model page for
+its verified fine-tuned checkpoint link. Select a final fine-tuned repository
 and enable REQUIRE_FINETUNED_SLICE to enforce post-LoRA evaluation; this guard
 refuses a baseline without finetune_manifest.json. Baseline SliceGPT scores
 must not be presented as fine-tuned scores.
@@ -80,7 +83,10 @@ https://github.com/EleutherAI/lm-evaluation-harness
 https://huggingface.co/docs/accelerate/concept_guides/big_model_inference
 
 The companion kaggle_slicegpt_lora.ipynb completes the nine unfinished sequential
-fine-tunes using the same diffusion objective and LoRA hyperparameters. Choose
+fine-tunes using the same diffusion objective and LoRA hyperparameters.
+Base checkpoints train sequentially on WikiText-2 → C4 (2048 optimizer steps);
+Instruct checkpoints add Alpaca (3072 steps). Each stage uses 4096 examples,
+2048 tokens, batch size 1 and gradient accumulation 4. Choose
 one original SliceGPT source per session, use both T4 GPUs or a larger GPU, and
 keep all weights on GPUs. The trainer refuses CPU/disk training offload.
 FP16 training is explicitly recorded as a separate precision profile; matching

bundle_manifest.json

@@ -1,6 +1,6 @@
 {
   "run_id": "run-9fc515e82f313fab",
-  "zip_sha256": "89b29760ba2b7ab074568c6480277a7e859b9b80bd1f8dc3f0d9afba6977e123",
+  "zip_sha256": "6097ccf56cdc81066d600c7ab7381b210a5b9fddeb553216d3a0e142819bb3f4",
   "available_checkpoints": 27,
   "shared_config": "Pruning-LLMs/diffusion_eval_and_vis/config.py",
   "smoke_tests": "All three methods on tiny LLaDA/Dream: likelihood, generation, CPU/disk and CUDA fp16 offload, reload agreement; task sharding and merging"

finetune_runner.py

@@ -53,7 +53,8 @@ def main():
     cfg.eval.dtype, cfg.eval.device = args.dtype, args.device
     if args.show_config:
         print(json.dumps({"finetune_config": asdict(cfg.finetune), "eval_config": asdict(cfg.eval),
-                          "dataset_revisions": DATASET_REVISIONS}, indent=2))
+                          "dataset_revisions": DATASET_REVISIONS,
+                          "effective_dataset_order": cfg.finetune.datasets_for_model(args.repo)}, indent=2))
         return
     import torch
     from huggingface_hub import HfApi, hf_hub_download

kaggle_slicegpt_lora.ipynb

@@ -17,7 +17,7 @@
       "metadata": {},
       "source": [
         "# Sequential SliceGPT LoRA on Kaggle\n",
-        "This is the same masked-diffusion LoRA procedure as the stopped Modal training: one adapter, WikiText-2 \u2192 C4 \u2192 Alpaca, pinned training datasets, rank 8/alpha 32, 4096 examples per stage, 2048 tokens, accumulation 4 and 3072 optimizer steps total.\n",
+        "This is the same masked-diffusion LoRA procedure as Modal training: one adapter and pinned training datasets. Base models use WikiText-2 \u2192 C4 (2048 optimizer steps). Instruct models use WikiText-2 \u2192 C4 \u2192 Alpaca (3072 steps). All use rank 8/alpha 32, 4096 examples per stage, 2048 tokens and accumulation 4. The setup cell prints the effective dataset order and full shared configuration.\n",
         "\n",
         "Enable Internet and GPU. On T4 choose both GPUs; the trainer requires all weights on GPUs, with no CPU/disk training offload. FP16 is an explicit precision change; BF16 is available on supporting GPUs. Hyperparameters stay unchanged. Every 50 steps and stage boundary saves optimizer, adapter and RNG state. A session time budget saves and exits so another session can continue.\n"
       ],
@@ -61,6 +61,7 @@
         "    print(i, row[\"repo_id\"])\n",
         "MODEL_INDEX = 0\n",
         "REPO = slice_models[MODEL_INDEX][\"repo_id\"]\n",
+        "REVISION = slice_models[MODEL_INDEX][\"revision\"]  # Stable source SHA for exact resume.\n",
         "DTYPE = \"float16\"                     # Explicit T4 profile; choose bfloat16 for matching supported hardware.\n",
         "SESSION_HOURS = 8\n",
         "TRAIN_OUTPUT = WORK_DIR / \"training_results\"\n",
@@ -90,7 +91,7 @@
         "        from getpass import getpass\n",
         "        os.environ[\"HF_TOKEN\"] = getpass(\"Hugging Face write token: \")\n",
         "TRAIN = [sys.executable, str(EVAL_ROOT / \"finetune_runner.py\"),\n",
-        "         \"--repo\", REPO, \"--dtype\", DTYPE, \"--output\", str(TRAIN_OUTPUT),\n",
+        "         \"--repo\", REPO, \"--revision\", REVISION, \"--dtype\", DTYPE, \"--output\", str(TRAIN_OUTPUT),\n",
         "         \"--cache-dir\", str(TEMP_DIR / \"structured_training_hf\"),\n",
         "         \"--offload-dir\", str(TEMP_DIR / \"structured_training_offload\"),\n",
         "         \"--time-budget-seconds\", str(SESSION_HOURS * 3600), \"--namespace\", HF_NAMESPACE]\n",

portable_eval.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:89b29760ba2b7ab074568c6480277a7e859b9b80bd1f8dc3f0d9afba6977e123
-size 66228
+oid sha256:6097ccf56cdc81066d600c7ab7381b210a5b9fddeb553216d3a0e142819bb3f4
+size 67178