Upload folder using huggingface_hub

frontend/src/components/RunWithLlmPane.tsx

@@ -718,7 +718,7 @@ function ModelPresetPicker({
             key={preset.id}
             preset={preset}
             selected={selected.id === preset.id}
-            disabled={disabled}
+            disabled={disabled ?? false}
             onSelect={() => selectPreset(preset)}
           />
         ))}

train/physix_train_colab.ipynb

@@ -323,28 +323,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "import subprocess, sys\n",
-    "\n",
-    "SFT_OUT = \"/tmp/physix-sft\"\n",
-    "\n",
-    "sft_cmd = [\n",
-    "    sys.executable, \"-m\", \"physix.training.sft\",\n",
-    "    \"--model\", p[\"base_model\"],\n",
-    "    \"--output-dir\", SFT_OUT,\n",
-    "    \"--epochs\", p[\"sft_epochs\"],\n",
-    "    \"--instances-per-system\", str(INSTANCES_PER_SYSTEM),\n",
-    "    \"--lora-r\", p[\"sft_lora_r\"],\n",
-    "    \"--learning-rate\", p[\"sft_lr\"],\n",
-    "    \"--wandb-run-name\", p[\"sft_run_name\"],\n",
-    "    \"--seed\", \"0\",\n",
-    "]\n",
-    "if SYSTEM_ID:\n",
-    "    sft_cmd += [\"--system-ids\", SYSTEM_ID]\n",
-    "if PUSH_TO_HUB and os.environ.get(\"HF_TOKEN\"):\n",
-    "    sft_cmd += [\"--hub-checkpoint-repo-id\", p[\"hub_ckpt_repo\"]]\n",
-    "\n",
-    "print(\"$\", \" \".join(sft_cmd))\n",
-    "subprocess.run(sft_cmd, check=True)"
+    "import subprocess, sys\n\nSFT_OUT = \"/tmp/physix-sft\"\n\nsft_cmd = [\n    sys.executable, \"-m\", \"physix.training.sft\",\n    \"--model\", p[\"base_model\"],\n    \"--output-dir\", SFT_OUT,\n    \"--epochs\", p[\"sft_epochs\"],\n    \"--instances-per-system\", str(INSTANCES_PER_SYSTEM),\n    \"--lora-r\", p[\"sft_lora_r\"],\n    \"--learning-rate\", p[\"sft_lr\"],\n    \"--wandb-run-name\", p[\"sft_run_name\"],\n    \"--seed\", \"0\",\n]\nif SYSTEM_ID:\n    sft_cmd += [\"--system-ids\", SYSTEM_ID]\nif PUSH_TO_HUB and os.environ.get(\"HF_TOKEN\"):\n    sft_cmd += [\"--hub-checkpoint-repo-id\", p[\"hub_ckpt_repo\"]]\n\nprint(\"=\" * 78)\nprint(\"  SFT is launched as a subprocess. Per-step loss WILL NOT stream to this cell.\")\nprint(\"  \u2192 Live loss curve:  https://wandb.ai/pratyush01/physix-live\")\nprint(f\"     run name:        {p['sft_run_name']}\")\nprint(\"  (If WANDB_API_KEY is unset, metrics are still saved to trainer_state.json\")\nprint(\"   under SFT_OUT and plotted in the final cell of this notebook.)\")\nprint(\"=\" * 78)\nprint(\"$\", \" \".join(sft_cmd))\nsubprocess.run(sft_cmd, check=True)\n"
    ]
   },
   {
@@ -372,35 +351,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "GRPO_OUT = \"/tmp/physix-grpo\"\n",
-    "\n",
-    "grpo_cmd = [\n",
-    "    sys.executable, \"-m\", \"physix.training.loop\",\n",
-    "    \"--model\", p[\"base_model\"],\n",
-    "    \"--output-dir\", GRPO_OUT,\n",
-    "    \"--num-steps\", p[\"num_steps\"],\n",
-    "    \"--num-generations\", p[\"num_generations\"],\n",
-    "    \"--max-completion-length\", p[\"max_completion\"],\n",
-    "    \"--learning-rate\", p[\"grpo_lr\"],\n",
-    "    \"--instances-per-system\", str(INSTANCES_PER_SYSTEM),\n",
-    "    \"--lora-r\", p[\"grpo_lora_r\"],\n",
-    "    \"--save-method\", \"merged_16bit\",\n",
-    "    \"--wandb-project\", \"physix-live\",\n",
-    "    \"--wandb-run-name\", p[\"grpo_run_name\"],\n",
-    "    \"--sft-checkpoint\", f\"{SFT_OUT}/merged\",\n",
-    "    \"--seed\", \"0\",\n",
-    "]\n",
-    "if SYSTEM_ID:\n",
-    "    grpo_cmd += [\"--system-ids\", SYSTEM_ID]\n",
-    "if PUSH_TO_HUB and os.environ.get(\"HF_TOKEN\"):\n",
-    "    grpo_cmd += [\n",
-    "        \"--push-to-hub\",\n",
-    "        \"--hub-repo-id\", p[\"hub_final_repo\"],\n",
-    "        \"--hub-checkpoint-repo-id\", p[\"hub_ckpt_repo\"],\n",
-    "    ]\n",
-    "\n",
-    "print(\"$\", \" \".join(grpo_cmd))\n",
-    "subprocess.run(grpo_cmd, check=True)"
+    "GRPO_OUT = \"/tmp/physix-grpo\"\n\ngrpo_cmd = [\n    sys.executable, \"-m\", \"physix.training.loop\",\n    \"--model\", p[\"base_model\"],\n    \"--output-dir\", GRPO_OUT,\n    \"--num-steps\", p[\"num_steps\"],\n    \"--num-generations\", p[\"num_generations\"],\n    \"--max-completion-length\", p[\"max_completion\"],\n    \"--learning-rate\", p[\"grpo_lr\"],\n    \"--instances-per-system\", str(INSTANCES_PER_SYSTEM),\n    \"--lora-r\", p[\"grpo_lora_r\"],\n    \"--save-method\", \"merged_16bit\",\n    \"--wandb-project\", \"physix-live\",\n    \"--wandb-run-name\", p[\"grpo_run_name\"],\n    \"--sft-checkpoint\", f\"{SFT_OUT}/merged\",\n    \"--seed\", \"0\",\n]\nif SYSTEM_ID:\n    grpo_cmd += [\"--system-ids\", SYSTEM_ID]\nif PUSH_TO_HUB and os.environ.get(\"HF_TOKEN\"):\n    grpo_cmd += [\n        \"--push-to-hub\",\n        \"--hub-repo-id\", p[\"hub_final_repo\"],\n        \"--hub-checkpoint-repo-id\", p[\"hub_ckpt_repo\"],\n    ]\n\nprint(\"=\" * 78)\nprint(\"  GRPO is launched as a subprocess. Per-step reward / loss WILL NOT stream here.\")\nprint(\"  \u2192 Live curves (reward, reward_std, KL, per-component reward, loss):\")\nprint(\"       https://wandb.ai/pratyush01/physix-live\")\nprint(f\"     run name: {p['grpo_run_name']}\")\nprint(\"  (If WANDB_API_KEY is unset, all metrics are still saved to\")\nprint(\"   trainer_state.json under GRPO_OUT and plotted in the final cell.)\")\nprint(\"=\" * 78)\nprint(\"$\", \" \".join(grpo_cmd))\nsubprocess.run(grpo_cmd, check=True)\n"
    ]
   },
   {