Upload prune-lora-v2.ipynb

prune-lora-v2.ipynb

@@ -0,0 +1 @@
+{"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.12.12"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceType":"datasetVersion","sourceId":16415322,"datasetId":8022630,"databundleVersionId":17411621,"mountSlug":"datasets/nekos4lyfe/image-caption-dataset"}],"dockerImageVersionId":31329,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true},"papermill":{"default_parameters":{},"duration":11372.148303,"end_time":"2026-05-14T21:19:13.559262+00:00","environment_variables":{},"exception":null,"input_path":"__notebook__.ipynb","output_path":"__notebook__.ipynb","parameters":{},"start_time":"2026-05-14T18:09:41.410959+00:00","version":"2.7.0"},"colab":{"provenance":[]}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# =============================================================================\n# CELL 1: Install dependencies\n# =============================================================================\n!pip install -q safetensors torch matplotlib numpy huggingface_hub pynacl py7zr\n\nimport os\nimport torch\nimport gc\nimport random\nimport datetime\n\nprint(f\"✅ PyTorch version: {torch.__version__}\")\nprint(f\"✅ Available GPUs: {torch.cuda.device_count()}\")\nfor i in range(torch.cuda.device_count()):\n    print(f\"   GPU {i}: {torch.cuda.get_device_name(i)}\")","metadata":{"execution":{"iopub.execute_input":"2026-05-14T18:09:45.284673Z","iopub.status.busy":"2026-05-14T18:09:45.284222Z","iopub.status.idle":"2026-05-14T18:10:01.729025Z","shell.execute_reply":"2026-05-14T18:10:01.727975Z"},"papermill":{"duration":16.450904,"end_time":"2026-05-14T18:10:01.730955+00:00","exception":false,"start_time":"2026-05-14T18:09:45.280051+00:00","status":"completed"},"tags":[],"id":"507c1488"},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =============================================================================\n# CELL 2: Secrets + Config from Kaggle Dataset\n# =============================================================================\nfrom kaggle_secrets import UserSecretsClient\nfrom huggingface_hub import snapshot_download, login\nimport torch\nimport safetensors.torch\nfrom safetensors import safe_open\nimport os\nimport random\nimport datetime\nimport gc\nfrom pathlib import Path\nfrom concurrent.futures import ThreadPoolExecutor, as_completed\nfrom collections import defaultdict\n\nprint(\"✅ Libraries imported!\")\n\n# === Kaggle Dataset Path ===\nkaggle_repo_path = \"/kaggle/input/datasets/nekos4lyfe/image-caption-dataset/\"  #@param {type:\"string\"}\n\n# Read HF_TOKEN from HF_TOKEN.txt\nhf_token = None\ntry:\n    token_path = Path(kaggle_repo_path) / \"HF_TOKEN.txt\"\n    if token_path.exists():\n        hf_token = token_path.read_text(encoding=\"utf-8\").strip()\n        print(\"✅ HF_TOKEN loaded from HF_TOKEN.txt\")\n\n        if hf_token:\n            login(token=hf_token)\n            print(\"✅ Successfully logged into Hugging Face\")\n    else:\n        print(f\"⚠️ HF_TOKEN.txt not found at {token_path}\")\nexcept Exception as e:\n    print(f\"⚠️ Failed to read HF_TOKEN.txt: {e}\")","metadata":{"execution":{"iopub.execute_input":"2026-05-14T18:10:01.737730Z","iopub.status.busy":"2026-05-14T18:10:01.736685Z","iopub.status.idle":"2026-05-14T18:10:03.127601Z","shell.execute_reply":"2026-05-14T18:10:03.126560Z"},"papermill":{"duration":1.395897,"end_time":"2026-05-14T18:10:03.129339+00:00","exception":false,"start_time":"2026-05-14T18:10:01.733442+00:00","status":"completed"},"tags":[],"id":"fc6693b0"},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =============================================================================\n# CELL 3: Setup Folders\n# =============================================================================\nbase_dir = \"/kaggle/working/pruned_loras\"\nos.makedirs(base_dir, exist_ok=True)\nos.makedirs(f\"{base_dir}/processed\", exist_ok=True)\nos.makedirs(f\"{base_dir}/metadata\", exist_ok=True)\n\nprint(f\"Working directory: {base_dir}\")\n\n# Read repo_id from repo_id.txt\nrepo_id = None\ntry:\n    repo_id_path = Path(kaggle_repo_path) / \"repo_id.txt\"\n    if repo_id_path.exists():\n        repo_id = repo_id_path.read_text(encoding=\"utf-8\").strip()\n        print(\"✅ repo_id loaded from repo_id.txt\")\n    else:\n        print(f\"⚠️ repo_id.txt not found at {repo_id_path}\")\nexcept Exception as e:\n    print(f\"⚠️ Failed to read repo_id.txt: {e}\")\n#-----#","metadata":{"execution":{"iopub.execute_input":"2026-05-14T18:10:03.140540Z","iopub.status.busy":"2026-05-14T18:10:03.139445Z","iopub.status.idle":"2026-05-14T18:10:03.153972Z","shell.execute_reply":"2026-05-14T18:10:03.152857Z"},"papermill":{"duration":0.019317,"end_time":"2026-05-14T18:10:03.155565+00:00","exception":false,"start_time":"2026-05-14T18:10:03.136248+00:00","status":"completed"},"tags":[],"id":"d3cddbf0"},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =============================================================================\n# CELL 4: Download + Smart Letter Naming\n# =============================================================================\n\nrepo_id = \"codeShare/flux-klein-9B-loras\"  #@param {type:\"string\"}\n\nprint(f\"Downloading .safetensors from {repo_id} ...\")\n\n# Updated allowed patterns\nsnapshot_download(\n    repo_id=repo_id,\n    local_dir=base_dir,\n    allow_patterns=[\n        'unprocessed*.safetensors',\n    ],\n    token=hf_token,\n)\n\noriginal_files = sorted([f for f in os.listdir(base_dir) if f.endswith(\".safetensors\")])\nprint(f\"✅ Downloaded {len(original_files)} LoRA files.\")\n\n# ====================== SMART LETTER GENERATOR ======================\ndef generate_letter_labels(n):\n    labels = []\n    i = 0\n    while len(labels) < n:\n        if i < 26:\n            labels.append(chr(65 + i))\n        else:\n            first = chr(65 + (i - 26) // 26)\n            second = chr(65 + (i - 26) % 26)\n            labels.append(first + second)\n        i += 1\n    return labels\n\nletters = generate_letter_labels(len(original_files))\n\n# Create mapping file BEFORE renaming\nmapping_path = os.path.join(base_dir, \"original_mapping.txt\")\nwith open(mapping_path, \"w\", encoding=\"utf-8\") as f:\n    f.write(\"Original Name → New Short Name\\n\")\n    f.write(\"=\" * 70 + \"\\n\")\n    for old_name, letter in zip(original_files, letters):\n        f.write(f\"{old_name}  →  {letter}.safetensors\\n\")\n\nprint(f\"✅ Mapping saved to: {mapping_path}\")\n\n# Rename files\nlora_files = []\nfor old_name, letter in zip(original_files, letters):\n    new_name = f\"{letter}.safetensors\"\n    old_path = os.path.join(base_dir, old_name)\n    new_path = os.path.join(base_dir, new_name)\n\n    os.rename(old_path, new_path)\n    lora_files.append(new_name)\n    print(f\"   {old_name}  →  {new_name}\")\n\nprint(f\"✅ All files renamed using A-Z + AA, AB... system.\")","metadata":{"execution":{"iopub.execute_input":"2026-05-14T18:10:03.161814Z","iopub.status.busy":"2026-05-14T18:10:03.161246Z","iopub.status.idle":"2026-05-14T18:10:26.771985Z","shell.execute_reply":"2026-05-14T18:10:26.770854Z"},"papermill":{"duration":23.615718,"end_time":"2026-05-14T18:10:26.773639+00:00","exception":false,"start_time":"2026-05-14T18:10:03.157921+00:00","status":"completed"},"tags":[],"id":"d3005a43"},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =============================================================================\n# CELL 5: Pruning - Supports Regular LoRA + LoKr (Kronecker)\n# =============================================================================\nfrom collections import defaultdict\nimport time\nimport datetime\nimport os\nimport random\nimport gc\nimport torch\nfrom concurrent.futures import ThreadPoolExecutor, as_completed\nfrom pathlib import Path\nfrom safetensors import safe_open\n\ndesired_accuracy = 95  #@param {type:\"slider\", min:80, max:100, step:1}\n\nprint(f\"🎯 Target Accuracy: {desired_accuracy}% | LoRA + LoKr Support\")\n\n# ====================== LOAD ORIGINAL NAME MAPPING ======================\noriginal_to_letter = {}\nmapping_path = os.path.join(base_dir, \"original_mapping.txt\")\nif os.path.exists(mapping_path):\n    with open(mapping_path, \"r\", encoding=\"utf-8\") as f:\n        for line in f:\n            if \"→\" in line:\n                parts = line.strip().split(\"→\")\n                if len(parts) == 2:\n                    old = parts[0].strip()\n                    new = parts[1].strip()\n                    original_to_letter[new] = old\n    print(f\"✅ Loaded {len(original_to_letter)} original filename mappings\")\nelse:\n    print(\"⚠️ original_mapping.txt not found - using short names\")\n\n# ====================== WORKER FUNCTION ======================\ndef worker(thread_id: int, gpu_id: int, file_list: list):\n    print(f\"[Thread {thread_id} | GPU {gpu_id}] 🚀 Started with {len(file_list)} files\")\n    torch.cuda.set_device(gpu_id)\n\n    for idx, filename in enumerate(file_list):\n        letter = filename.replace(\".safetensors\", \"\")\n        input_path = os.path.join(base_dir, filename)\n        size_before = os.path.getsize(input_path) / (1024 * 1024)\n\n        print(f\"[Thread {thread_id} | GPU {gpu_id}] 🔄 Processing [{letter}] {filename} ({size_before:.2f} MB)\")\n\n        torch.cuda.empty_cache()\n        gc.collect()\n\n        original_filename = original_to_letter.get(filename, filename)\n\n        random_suffix = random.randint(10000, 99999)\n        output_name = f\"{letter}_{random_suffix}.safetensors\"\n        output_path = f\"{base_dir}/processed/{output_name}\"\n\n        # Load file\n        state_dict = safetensors.torch.load_file(input_path, device=\"cpu\")\n\n        original_metadata = {}\n        try:\n            with safe_open(input_path, framework=\"pt\", device=\"cpu\") as f:\n                original_metadata = f.metadata() or {}\n        except:\n            pass\n\n        # ====================== PARSE LAYERS (LoRA + LoKr) ======================\n        lora_pairs = defaultdict(dict)\n        original_ranks = {}\n        is_lokr = False\n\n        for key, tensor in state_dict.items():\n            if \".lokr_\" in key:\n                is_lokr = True\n                # Extract base key\n                base = key.split(\".lokr_\")[0]\n\n                if \".lokr_w1\" in key:\n                    lora_pairs[base][\"w1\"] = tensor\n                    original_ranks[base] = tensor.shape[0] * tensor.shape[1]  # approximate full rank\n                elif \".lokr_w2\" in key:\n                    lora_pairs[base][\"w2\"] = tensor\n                elif \".lokr_w\" in key:   # some LoKr have single lokr_w\n                    lora_pairs[base][\"w\"] = tensor\n                    original_ranks[base] = tensor.shape[0]\n\n            elif \".lora_A.weight\" in key or \".lora_down.weight\" in key:\n                base = key.replace(\".lora_A.weight\", \"\").replace(\".lora_down.weight\", \"\")\n                lora_pairs[base][\"A\"] = tensor\n                original_ranks[base] = tensor.shape[0]\n            elif \".lora_B.weight\" in key or \".lora_up.weight\" in key:\n                base = key.replace(\".lora_B.weight\", \"\").replace(\".lora_up.weight\", \"\")\n                lora_pairs[base][\"B\"] = tensor\n\n        # Pruning\n        new_state_dict = dict(state_dict)\n        changes_made = False\n        pruned_count = 0\n        total_layers = len(lora_pairs)\n        layer_log = []\n\n        for base_key, matrices in lora_pairs.items():\n            if not matrices:\n                continue\n\n            # ==================== LoKr Handling ====================\n            if \"w1\" in matrices and \"w2\" in matrices:\n                # Reconstruct delta_W = kron(w1, w2)\n                w1 = matrices[\"w1\"].to(gpu_id)\n                w2 = matrices[\"w2\"].to(gpu_id)\n                delta_W = torch.kron(w1, w2).float()\n                orig_rank = original_ranks.get(base_key, w1.shape[0] * w1.shape[1])\n                is_lokr_layer = True\n            elif \"w\" in matrices:\n                delta_W = matrices[\"w\"].to(gpu_id).float()\n                orig_rank = original_ranks.get(base_key, delta_W.shape[0])\n                is_lokr_layer = True\n            # ==================== Standard LoRA ====================\n            elif \"A\" in matrices and \"B\" in matrices:\n                A = matrices[\"A\"].to(gpu_id)\n                B = matrices[\"B\"].to(gpu_id)\n                delta_W = torch.matmul(B, A).float()\n                orig_rank = original_ranks.get(base_key, A.shape[0])\n                is_lokr_layer = False\n            else:\n                continue\n\n            original_norm = torch.norm(delta_W, p='fro').item()\n\n            U, S, Vh = torch.linalg.svd(delta_W, full_matrices=False)\n\n            # Try reduced ranks\n            candidates = [max(1, int(orig_rank * f)) for f in [0.95, 0.9, 0.85, 0.8, 0.75, 0.7, 0.65, 0.6]]\n            test_ranks = sorted(set(r for r in candidates if r < orig_rank))\n\n            best_rank = orig_rank\n            best_acc = 100.0\n\n            for r in test_ranks:\n                if r >= len(S): continue\n                S_trunc = torch.zeros_like(S)\n                S_trunc[:r] = S[:r]\n                approx = U @ torch.diag(S_trunc) @ Vh\n                error = torch.norm(delta_W - approx, p='fro').item()\n                rel_error = (error / original_norm * 100) if original_norm > 1e-8 else 0.0\n                acc = 100 - rel_error\n\n                if acc >= desired_accuracy:\n                    best_rank = r\n                    best_acc = acc\n                    break\n\n            if best_rank < orig_rank:\n                changes_made = True\n                pruned_count += 1\n                r = best_rank\n\n                sqrt_S = torch.sqrt(S[:r])\n                new_up = (U[:, :r] * sqrt_S.unsqueeze(0)).to(torch.bfloat16)\n                new_down = (sqrt_S.unsqueeze(1) * Vh[:r, :]).to(torch.bfloat16)\n\n                # Remove ALL old keys for this layer\n                for suffix in [\".lora_A.weight\", \".lora_B.weight\",\n                              \".lora_down.weight\", \".lora_up.weight\",\n                              \".lokr_w1\", \".lokr_w2\", \".lokr_w\", \".alpha\"]:\n                    new_state_dict.pop(base_key + suffix, None)\n\n                # Save as standard LoRA (most compatible)\n                new_state_dict[base_key + \".lora_down.weight\"] = new_down.contiguous().cpu()\n                new_state_dict[base_key + \".lora_up.weight\"]   = new_up.contiguous().cpu()\n\n                print(f\"   → {base_key} | {orig_rank} → {best_rank} | Acc {best_acc:.2f}% {'(LoKr)' if is_lokr_layer else ''}\")\n                layer_log.append(f\"{base_key} | {orig_rank:3} → {best_rank:3} | Pruned | Acc {best_acc:.2f}% {'(LoKr)' if is_lokr_layer else ''}\")\n            else:\n                layer_log.append(f\"{base_key} | {orig_rank:3} → {orig_rank:3} | No change\")\n\n            # Cleanup\n            del delta_W, U, S, Vh\n            if 'w1' in locals(): del w1\n            if 'w2' in locals(): del w2\n            torch.cuda.empty_cache()\n\n        # Delete original\n        if os.path.exists(input_path):\n            os.remove(input_path)\n\n        if not changes_made:\n            print(f\"[Thread {thread_id} | GPU {gpu_id}] → No pruning needed for [{letter}]\")\n            continue\n\n        # ====================== SAVE FILE ======================\n        new_metadata = dict(original_metadata)\n        new_metadata.update({\n            \"original_file\": original_filename,\n            \"pruned_with\": f\"SVD Rank Reduction at {desired_accuracy}% accuracy\",\n            \"pruning_date\": datetime.datetime.now().isoformat(),\n            \"random_id\": str(random_suffix),\n            \"layers_pruned\": f\"{pruned_count}/{total_layers}\",\n            \"note\": \"Converted LoKr to standard LoRA where applicable\"\n        })\n\n        for k, v in new_state_dict.items():\n            if isinstance(v, torch.Tensor):\n                new_state_dict[k] = v.contiguous()\n\n        safetensors.torch.save_file(new_state_dict, output_path, metadata=new_metadata)\n\n        size_after = os.path.getsize(output_path) / (1024 * 1024)\n\n        # Save metadata log\n        meta_dir = f\"{base_dir}/metadata/{letter}\"\n        os.makedirs(meta_dir, exist_ok=True)\n        with open(f\"{meta_dir}/{letter}_metadata.txt\", \"w\", encoding=\"utf-8\") as f:\n            f.write(f\"Original filename: {original_filename}\\n\")\n            f.write(f\"Pruned filename : {output_name}\\n\")\n            f.write(f\"Size Before     : {size_before:.2f} MB\\n\")\n            f.write(f\"Size After      : {size_after:.2f} MB\\n\")\n            f.write(f\"Layers Pruned   : {pruned_count}/{total_layers}\\n\\n\")\n            f.write(\"=== LAYER DETAILS ===\\n\")\n            for line in layer_log:\n                f.write(line + \"\\n\")\n\n        print(f\"[Thread {thread_id} | GPU {gpu_id}] ✅ Saved [{letter}] | {size_before:.1f}MB → {size_after:.1f}MB\")\n\n        del state_dict, new_state_dict\n        torch.cuda.empty_cache()\n        gc.collect()\n\n    print(f\"[Thread {thread_id} | GPU {gpu_id}] 🏁 Worker completed\")\n    return True\n\n\n# ====================== LAUNCH THREADS ======================\nprint(f\"\\nDistributing {len(lora_files)} files...\")\n\nnum_threads = 2\nchunks = [[] for _ in range(num_threads)]\nfor i, f in enumerate(lora_files):\n    chunks[i % num_threads].append(f)\n\nwith ThreadPoolExecutor(max_workers=num_threads) as executor:\n    futures = []\n    for t in range(num_threads):\n        gpu_id = t % 2\n        future = executor.submit(worker, t, gpu_id, chunks[t])\n        futures.append(future)\n        print(f\"   → Thread {t} on GPU {gpu_id} ({len(chunks[t])} files)\")\n        if t < num_threads - 1:\n            time.sleep(15)\n\n    for future in as_completed(futures):\n        future.result()\n\nprint(\"\\n🎉 All pruning completed! (LoRA + LoKr supported)\")","metadata":{"_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","execution":{"iopub.execute_input":"2026-05-14T18:10:26.781049Z","iopub.status.busy":"2026-05-14T18:10:26.780630Z","iopub.status.idle":"2026-05-14T21:19:09.863745Z","shell.execute_reply":"2026-05-14T21:19:09.862844Z"},"papermill":{"duration":11323.149894,"end_time":"2026-05-14T21:19:09.926355+00:00","exception":false,"start_time":"2026-05-14T18:10:26.776461+00:00","status":"completed"},"tags":[],"id":"d328be3b"},"outputs":[],"execution_count":null}]}