Upload Create_SDNQ_from_klein_9b_transformer_model.ipynb

colab_notebooks/Create_SDNQ_from_klein_9b_transformer_model.ipynb

@@ -44,22 +44,10 @@
         "print(\"✅ Dependencies installed\")"
       ],
       "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "XXld_BQ1sU0v",
-        "outputId": "38508641-2a6a-4a19-f585-d2ac2d2faef4"
+        "id": "XXld_BQ1sU0v"
       },
-      "execution_count": 7,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "✅ Dependencies installed\n"
-          ]
-        }
-      ]
+      "execution_count": null,
+      "outputs": []
     },
     {
       "cell_type": "code",
@@ -128,7 +116,7 @@
       "metadata": {
         "id": "rQ5cr5ZsUQqG"
       },
-      "execution_count": 8,
+      "execution_count": 4,
       "outputs": []
     },
     {
@@ -239,9 +227,9 @@
           "base_uri": "https://localhost:8080/"
         },
         "id": "aGH03dqUQcsQ",
-        "outputId": "b5d8cd52-aa9b-49f6-c43b-0f9c75fe210f"
+        "outputId": "52fdb852-4c99-4e2c-f13e-7bfe0e62f511"
       },
-      "execution_count": 10,
+      "execution_count": 6,
       "outputs": [
         {
           "output_type": "stream",
@@ -265,7 +253,7 @@
         "id": "EpL4Kek7AbjZ",
         "outputId": "9c46f98d-cbd0-498f-bb94-33b6860f49c6"
       },
-      "execution_count": 7,
+      "execution_count": null,
       "outputs": [
         {
           "output_type": "execute_result",
@@ -551,31 +539,39 @@
         "pipe.transformer = sdnq_post_load_quant(\n",
         "    pipe.transformer,\n",
         "    use_dynamic_quantization=True,\n",
-        "    weights_dtype=\"uint4\",\n",
-        "    dynamic_loss_threshold=1e-2,\n",
+        "    weights_dtype=\"uint2\",\n",
+        "    dynamic_loss_threshold=1e-1,\n",
         "    use_svd=True,\n",
         "    group_size=0,\n",
+        "    svd_steps=16,\n",
         "    quantization_device=\"cuda\",\n",
         "    return_device=\"cpu\",\n",
         "    quant_conv=False,\n",
         "    quant_embedding=False,\n",
         ")\n",
         "\n",
-        "import torch\n",
+        "import torch , gc\n",
         "with torch.no_grad():\n",
         "  transformer_save_path = \"/content/transformer\"\n",
-        "  pipe.transformer.save_pretrained(transformer_save_path,safe_serialization=True , max_shard_size='2GB')\n",
-        "  print(f\"✅ pipe.transformer saved to: {transformer_save_path}\")"
+        "  pipe.transformer.save_pretrained(transformer_save_path,safe_serialization=True)\n",
+        "  print(f\"✅ pipe.transformer saved to: {transformer_save_path}\")\n",
+        "#----#\n",
+        "gc.collect()\n",
+        "torch.cuda.empty_cache()"
       ],
       "metadata": {
         "id": "hWAkx7u3XfdY"
       },
-      "execution_count": 12,
+      "execution_count": null,
       "outputs": []
     },
     {
       "cell_type": "code",
       "source": [
+        "import gc , torch\n",
+        "gc.collect()\n",
+        "torch.cuda.empty_cache()\n",
+        "\n",
         "import torch, gc\n",
         "from sdnq import SDNQConfig , sdnq_post_load_quant\n",
         "\n",
@@ -584,8 +580,8 @@
         "    pipe.text_encoder,\n",
         "    use_dynamic_quantization=True,\n",
         "    weights_dtype=\"uint2\",\n",
-        "    dynamic_loss_threshold=1e-2,\n",
-        "    use_svd=True,\n",
+        "    dynamic_loss_threshold=1e-1,\n",
+        "    use_svd=False,\n",
         "    group_size=0,\n",
         "    quantization_device=\"cuda\",\n",
         "    return_device=\"cpu\",\n",
@@ -596,13 +592,16 @@
         "import torch\n",
         "with torch.no_grad():\n",
         "  text_encoder_save_path = \"/content/text_encoder\"\n",
-        "  pipe.text_encoder.save_pretrained(text_encoder_save_path,safe_serialization=True , max_shard_size='2GB')\n",
-        "  print(f\"✅ pipe.text_encoder saved to: {text_encoder_save_path}\")"
+        "  pipe.text_encoder.save_pretrained(text_encoder_save_path,safe_serialization=True)\n",
+        "  print(f\"✅ pipe.text_encoder saved to: {text_encoder_save_path}\")\n",
+        "#----#\n",
+        "gc.collect()\n",
+        "torch.cuda.empty_cache()"
       ],
       "metadata": {
         "id": "uSPLPAVwgXp_"
       },
-      "execution_count": 10,
+      "execution_count": null,
       "outputs": []
     },
     {
@@ -615,7 +614,7 @@
       "metadata": {
         "id": "Xvnz0s2AV0VW"
       },
-      "execution_count": 3,
+      "execution_count": null,
       "outputs": []
     },
     {
@@ -627,7 +626,7 @@
         "import torch\n",
         "with torch.no_grad():\n",
         "  vae_save_path = \"/content/vae\"\n",
-        "  pipe.vae.save_pretrained(vae_save_path,safe_serialization=True , max_shard_size='2GB')\n",
+        "  pipe.vae.save_pretrained(vae_save_path,safe_serialization=True)\n",
         "  print(f\"✅ pipe.vae saved to: {vae_save_path}\")"
       ],
       "metadata": {
@@ -635,9 +634,9 @@
           "base_uri": "https://localhost:8080/"
         },
         "id": "4PBgpLyDlwn4",
-        "outputId": "b9419cc2-5b2d-4909-980d-c424b00e86e0"
+        "outputId": "eab87add-9978-45db-e585-ced9498ee6c8"
       },
-      "execution_count": 11,
+      "execution_count": 9,
       "outputs": [
         {
           "output_type": "stream",
@@ -654,7 +653,7 @@
         "import torch\n",
         "with torch.no_grad():\n",
         "  scheduler_save_path = \"/content/scheduler\"\n",
-        "  pipe.scheduler.save_pretrained(scheduler_save_path,safe_serialization=True , max_shard_size='2GB')\n",
+        "  pipe.scheduler.save_pretrained(scheduler_save_path,safe_serialization=True)\n",
         "  print(f\"✅ pipe.scheduler saved to: {scheduler_save_path}\")"
       ],
       "metadata": {
@@ -662,9 +661,9 @@
           "base_uri": "https://localhost:8080/"
         },
         "id": "nxJZeqwlmgcR",
-        "outputId": "e4155232-bd12-48fc-b80d-db1358c6f1fb"
+        "outputId": "1bfff9be-6236-4663-ff8a-ffb4394be809"
       },
-      "execution_count": 12,
+      "execution_count": 10,
       "outputs": [
         {
           "output_type": "stream",
@@ -681,7 +680,7 @@
         "import torch\n",
         "with torch.no_grad():\n",
         "  tokenizer_save_path = \"/content/tokenizer\"\n",
-        "  pipe.tokenizer.save_pretrained(tokenizer_save_path,safe_serialization=True , max_shard_size='2GB')\n",
+        "  pipe.tokenizer.save_pretrained(tokenizer_save_path,safe_serialization=True)\n",
         "  print(f\"✅ pipe.tokenizer saved to: {tokenizer_save_path}\")"
       ],
       "metadata": {
@@ -689,9 +688,9 @@
           "base_uri": "https://localhost:8080/"
         },
         "id": "ZGOf4WtqFH6g",
-        "outputId": "cac3ffff-efd6-4255-c7e5-86ae75921a00"
+        "outputId": "571daca1-dcf9-44a8-ecd7-9c460276964f"
       },
-      "execution_count": 13,
+      "execution_count": 11,
       "outputs": [
         {
           "output_type": "stream",
@@ -1021,8 +1020,7 @@
         "api = HfApi()\n",
         "\n",
         "# Define the target repository ID\n",
-        "repo_id = \"codeShare/FLUX.2-klein-9b-SDNQ-2bit\" # Or \"codeShare/FLUX.2-klein-9b-SDNQ-4bit\" if that was the intended repo\n",
-        "\n",
+        "repo_id = \"codeShare/FLUX.2-klein-9b-SDNQ-2bit\"\n",
         "# Check if repository exists, if not, create it\n",
         "print(f\"Checking if repository {repo_id} exists...\")\n",
         "try:\n",
@@ -1079,6 +1077,126 @@
       ],
       "execution_count": null,
       "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "f58c3a41"
+      },
+      "source": [
+        "import os\n",
+        "from huggingface_hub import HfApi, login\n",
+        "from google.colab import userdata\n",
+        "\n",
+        "hf_token = userdata.get(\"HF_TOKEN\")\n",
+        "login(token=hf_token)\n",
+        "\n",
+        "api = HfApi()\n",
+        "\n",
+        "repo_id = \"codeShare/FLUX.2-klein-9b-SDNQ-2bit\"\n",
+        "source_folder_path = \"/content/transformer\"\n",
+        "target_folder_in_repo = \"transformer\"\n",
+        "\n",
+        "print(f\"\\nUploading folder: {source_folder_path} to repo path: {target_folder_in_repo} in {repo_id}...\")\n",
+        "\n",
+        "if os.path.isdir(source_folder_path):\n",
+        "    api.upload_folder(\n",
+        "        folder_path=source_folder_path,\n",
+        "        repo_id=repo_id,\n",
+        "        repo_type=\"model\",\n",
+        "        path_in_repo=target_folder_in_repo,\n",
+        "        commit_message=f\"Upload {target_folder_in_repo} component\"\n",
+        "    )\n",
+        "    print(f\"✅ Folder {source_folder_path} uploaded to {repo_id}/{target_folder_in_repo}.\")\n",
+        "else:\n",
+        "    print(f\"⚠️ Source folder not found, skipping: {source_folder_path}\")\n",
+        "\n",
+        "print(\"\\n✅ Transformer component upload process complete.\")"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "ae55ccc9"
+      },
+      "source": [
+        "import os\n",
+        "from huggingface_hub import HfApi, login\n",
+        "from google.colab import userdata\n",
+        "\n",
+        "hf_token = userdata.get(\"HF_TOKEN\")\n",
+        "login(token=hf_token)\n",
+        "\n",
+        "api = HfApi()\n",
+        "\n",
+        "repo_id = \"codeShare/FLUX.2-klein-9b-SDNQ-2bit\"\n",
+        "source_folder_path = \"/content/text_encoder\"\n",
+        "target_folder_in_repo = \"text_encoder\"\n",
+        "\n",
+        "print(f\"\\nUploading folder: {source_folder_path} to repo path: {target_folder_in_repo} in {repo_id}...\")\n",
+        "\n",
+        "if os.path.isdir(source_folder_path):\n",
+        "    api.upload_folder(\n",
+        "        folder_path=source_folder_path,\n",
+        "        repo_id=repo_id,\n",
+        "        repo_type=\"model\",\n",
+        "        path_in_repo=target_folder_in_repo,\n",
+        "        commit_message=f\"Upload {target_folder_in_repo} component\"\n",
+        "    )\n",
+        "    print(f\"✅ Folder {source_folder_path} uploaded to {repo_id}/{target_folder_in_repo}.\")\n",
+        "else:\n",
+        "    print(f\"⚠️ Source folder not found, skipping: {source_folder_path}\")\n",
+        "\n",
+        "print(\"\\n✅ Text Encoder component upload process complete.\")"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "10131a92"
+      },
+      "source": [
+        "import os\n",
+        "from huggingface_hub import HfApi, login\n",
+        "from google.colab import userdata\n",
+        "\n",
+        "hf_token = userdata.get(\"HF_TOKEN\")\n",
+        "login(token=hf_token)\n",
+        "\n",
+        "api = HfApi()\n",
+        "\n",
+        "repo_id = \"codeShare/FLUX.2-klein-9b-SDNQ-2bit\"\n",
+        "\n",
+        "folders_to_upload = {\n",
+        "    \"/content/vae\": \"vae\",\n",
+        "    \"/content/scheduler\": \"scheduler\",\n",
+        "    \"/content/tokenizer\": \"tokenizer\",\n",
+        "}\n",
+        "\n",
+        "print(f\"\\nStarting upload of vae, scheduler, and tokenizer components to {repo_id}...\")\n",
+        "\n",
+        "for source_folder, target_folder_in_repo in folders_to_upload.items():\n",
+        "    print(f\"Uploading folder: {source_folder} to repo path: {target_folder_in_repo} in {repo_id}...\")\n",
+        "    if os.path.isdir(source_folder):\n",
+        "        api.upload_folder(\n",
+        "            folder_path=source_folder,\n",
+        "            repo_id=repo_id,\n",
+        "            repo_type=\"model\",\n",
+        "            path_in_repo=target_folder_in_repo,\n",
+        "            commit_message=f\"Upload {target_folder_in_repo} component\"\n",
+        "        )\n",
+        "        print(f\"✅ Folder {source_folder} uploaded to {repo_id}/{target_folder_in_repo}.\")\n",
+        "    else:\n",
+        "        print(f\"⚠️ Source folder not found, skipping: {source_folder}\")\n",
+        "\n",
+        "print(\"\\n✅ All requested components upload process complete.\")"
+      ],
+      "execution_count": null,
+      "outputs": []
     }
   ],
   "metadata": {