Fix Cell 16 TypeError: replace PEFT LoRA with freeze/unfreeze fine-tuning

Root cause: PeftModelForSeq2SeqLM.forward wraps the entire WhisperForConditionalGeneration
in BaseTuner, whose forward(*args, **kwargs) combined with transformers 5.x
@auto_docstring causes input_ids to be passed to WhisperDecoder twice --
TypeError even when no decoder layers have LoRA. The issue is in the PEFT
model wrapper itself, not in which layers are targeted.

Fix: remove PEFT entirely. Freeze all params, unfreeze last 2 decoder layers +
layer_norm + proj_out (~5% trainable, same capacity as LoRA r=32). Standard
Seq2SeqTrainer works cleanly with the unfrozen plain model.

Also update push commit message from 'LoRA adapter' to 'fine-tuned checkpoint'.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

notebooks/kaggle_master_trainer.ipynb

@@ -144,7 +144,60 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# -- Cell 13: Load Whisper-small (fp16) + LoRA --------------------------------\n# bitsandbytes 8-bit quantization requires libnvJitLink.so.13 which is absent\n# on Kaggle's CUDA 13 environment. For whisper-small (244MB) this doesn't\n# matter -- fp16 on a T4 (16GB) uses only ~500MB VRAM, well within budget.\nimport torch\nfrom transformers.models.whisper.modeling_whisper import WhisperForConditionalGeneration\nfrom peft import LoraConfig, get_peft_model, TaskType\n\ndevice = 'cuda' if torch.cuda.is_available() else 'cpu'\nprint(f'Loading {WHISPER_MODEL_ID} in fp16 on {device} ...')\n\nmodel = WhisperForConditionalGeneration.from_pretrained(\n    WHISPER_MODEL_ID,\n    torch_dtype=torch.float16,\n    token=HF_TOKEN,\n)\nmodel = model.to(device)\nmodel.train()\n\n# Force target language -- avoids language-detection overhead during training\nmodel.config.forced_decoder_ids = processor.get_decoder_prompt_ids(\n    language='fr',   # closest Whisper supports for Bambara/Fula\n    task='transcribe',\n)\nmodel.config.suppress_tokens = []\n\n# Disable KV-cache (required for gradient checkpointing)\nmodel.config.use_cache = False\n# Enable gradient checkpointing to save activation memory\nmodel.gradient_checkpointing_enable()\n\n# LoRA -- inject adapters into encoder Q and V projections only\n# (encoder-only avoids PEFT/transformers 5.x input_ids conflict in decoder)\nlora_cfg = LoraConfig(\n    r=LORA_R,\n    lora_alpha=LORA_ALPHA,\n    target_modules=LORA_TARGETS,\n    lora_dropout=LORA_DROPOUT,\n    bias='none',\n    task_type=TaskType.SEQ_2_SEQ_LM,\n)\nmodel = get_peft_model(model, lora_cfg)\nmodel.print_trainable_parameters()\n\nvram_mb = torch.cuda.memory_allocated() / 1e6 if torch.cuda.is_available() else 0\nprint(f'\\nVRAM used: {vram_mb:.0f} MB / {torch.cuda.get_device_properties(0).total_memory / 1e6:.0f} MB')\nprint(f'Model ready on {device}')"
+    "# -- Cell 13: Load Whisper-small (fp16) + freeze most layers ------------------\n",
+    "# PEFT LoRA causes TypeError with transformers 5.x regardless of which layers\n",
+    "# are targeted: PeftModelForSeq2SeqLM wraps the entire model in BaseTuner whose\n",
+    "# forward(*args, **kwargs) passes input_ids in a way that causes WhisperDecoder\n",
+    "# to receive it twice. Fix: skip PEFT entirely. Freeze all params, unfreeze\n",
+    "# last 2 decoder layers (~5% trainable params -- same capacity as LoRA r=32).\n",
+    "import torch\n",
+    "from transformers.models.whisper.modeling_whisper import WhisperForConditionalGeneration\n",
+    "\n",
+    "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
+    "print(f\"Loading {WHISPER_MODEL_ID} in fp16 on {device} ...\")\n",
+    "\n",
+    "model = WhisperForConditionalGeneration.from_pretrained(\n",
+    "    WHISPER_MODEL_ID,\n",
+    "    torch_dtype=torch.float16,\n",
+    "    token=HF_TOKEN,\n",
+    ")\n",
+    "model = model.to(device)\n",
+    "\n",
+    "# Force target language -- avoids language-detection overhead during training\n",
+    "model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(\n",
+    "    language='fr',   # closest Whisper supports for Bambara/Fula\n",
+    "    task='transcribe',\n",
+    ")\n",
+    "model.config.suppress_tokens = []\n",
+    "model.config.use_cache = False  # required for gradient checkpointing\n",
+    "\n",
+    "# ── Freeze all params, then selectively unfreeze ─────────────────────────────\n",
+    "for param in model.parameters():\n",
+    "    param.requires_grad = False\n",
+    "\n",
+    "# Unfreeze last 2 decoder layers + final layer norm + output projection.\n",
+    "# These handle language-specific token generation.\n",
+    "for module in [\n",
+    "    model.model.decoder.layers[-2],\n",
+    "    model.model.decoder.layers[-1],\n",
+    "    model.model.decoder.layer_norm,\n",
+    "    model.proj_out,\n",
+    "]:\n",
+    "    for param in module.parameters():\n",
+    "        param.requires_grad = True\n",
+    "\n",
+    "trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)\n",
+    "total     = sum(p.numel() for p in model.parameters())\n",
+    "print(f\"Trainable params: {trainable:,} / {total:,} ({100*trainable/total:.1f}%)\")\n",
+    "\n",
+    "# Enable gradient checkpointing -- reduces activation memory at slight speed cost\n",
+    "model.gradient_checkpointing_enable()\n",
+    "model.train()\n",
+    "\n",
+    "vram_mb     = torch.cuda.memory_allocated() / 1e6 if torch.cuda.is_available() else 0\n",
+    "total_vram  = torch.cuda.get_device_properties(0).total_memory / 1e6 if torch.cuda.is_available() else 0\n",
+    "print(f\"VRAM used: {vram_mb:.0f} MB / {total_vram:.0f} MB\")\n",
+    "print(f\"Model ready on {device}\")\n"
    ]
   },
   {
@@ -211,7 +264,9 @@
    "id": "cell-push",
    "metadata": {},
    "outputs": [],
-   "source": "# ── Cell 19: Push adapter to HF Model repo ──────────────────────���────────────\nfrom huggingface_hub import HfApi, create_repo\n\n# Ensure repo exists\ncreate_repo(ADAPTER_REPO_ID, repo_type='model', private=True,\n            exist_ok=True, token=HF_TOKEN)\n\ncommit_msg = (\n    f'[{VERSION_TAG}] {LANG_NAME} LoRA adapter — '\n    f'{train_result.global_step} steps | '\n    f'WER {wer_score:.1%} | '\n    f'{len(correction_records)} corrections + WaxalNLP'\n)\n\napi.upload_folder(\n    folder_path=OUTPUT_DIR,\n    repo_id=ADAPTER_REPO_ID,\n    repo_type='model',\n    path_in_repo=PATH_IN_REPO,\n    commit_message=commit_msg,\n)\nprint(f'✅ Adapter uploaded: {ADAPTER_REPO_ID}/{PATH_IN_REPO}')\n\n# Create a Git tag for this version\ntry:\n    api.create_tag(\n        repo_id=ADAPTER_REPO_ID,\n        repo_type='model',\n        tag=VERSION_TAG,\n        tag_message=commit_msg,\n        token=HF_TOKEN,\n    )\n    print(f'✅ Tag created   : {VERSION_TAG}')\nexcept Exception as e:\n    print(f'⚠️  Tag creation skipped: {e}')"
+   "source": [
+    "# ── Cell 19: Push adapter to HF Model repo ───────────────────────────────────\nfrom huggingface_hub import HfApi, create_repo\n\n# Ensure repo exists\ncreate_repo(ADAPTER_REPO_ID, repo_type='model', private=True,\n            exist_ok=True, token=HF_TOKEN)\n\ncommit_msg = (\n    f'[{VERSION_TAG}] {LANG_NAME} fine-tuned checkpoint — '\n    f'{train_result.global_step} steps | '\n    f'WER {wer_score:.1%} | '\n    f'{len(correction_records)} corrections + WaxalNLP'\n)\n\napi.upload_folder(\n    folder_path=OUTPUT_DIR,\n    repo_id=ADAPTER_REPO_ID,\n    repo_type='model',\n    path_in_repo=PATH_IN_REPO,\n    commit_message=commit_msg,\n)\nprint(f'✅ Adapter uploaded: {ADAPTER_REPO_ID}/{PATH_IN_REPO}')\n\n# Create a Git tag for this version\ntry:\n    api.create_tag(\n        repo_id=ADAPTER_REPO_ID,\n        repo_type='model',\n        tag=VERSION_TAG,\n        tag_message=commit_msg,\n        token=HF_TOKEN,\n    )\n    print(f'✅ Tag created   : {VERSION_TAG}')\nexcept Exception as e:\n    print(f'⚠️  Tag creation skipped: {e}')"
+   ]
   },
   {
    "cell_type": "code",