Adding naive `dist.all_gather` version for distributed computing

MuonForOLMo.ipynb

@@ -16,7 +16,7 @@
   "cells": [
     {
       "cell_type": "code",
-      "execution_count": 6,
+      "execution_count": 1,
       "metadata": {
         "id": "cCXb6F65XhI_"
       },
@@ -421,7 +421,7 @@
       "metadata": {
         "id": "o9dFXoh2YSVn"
       },
-      "execution_count": 7,
+      "execution_count": 2,
       "outputs": []
     },
     {
@@ -542,6 +542,8 @@
         "\n",
         "            for name, p in zip(group[\"param_names\"], group[\"params\"]):\n",
         "                name = self._clean_param_name(name)\n",
+        "                # Check if we're in FSDP mode\n",
+        "                is_fsdp = hasattr(p, '_is_sharded') and p._is_sharded\n",
         "\n",
         "                if p.grad is None:\n",
         "                    if collecting_metrics:\n",
@@ -582,20 +584,55 @@
         "                    if isinstance(mask, torch.Tensor):\n",
         "                        update.mul_(mask)\n",
         "\n",
-        "                    # Handle conv filters\n",
-        "                    orig_shape = update.shape\n",
-        "                    if update.ndim == 4:\n",
-        "                        update = update.view(update.shape[0], -1)\n",
         "\n",
-        "                    # Apply Newton-Schulz\n",
-        "                    update = self.zeropower_via_newtonschulz5(update, steps=ns_steps)\n",
+        "                    if is_fsdp:\n",
+        "                        # For FSDP, we need to gather the full gradient/update across ranks\n",
+        "                        import torch.distributed as dist\n",
         "\n",
-        "                    # Scale update\n",
-        "                    update *= max(1, grad.size(-2) / grad.size(-1)) ** 0.5\n",
+        "                        # Get world size and rank\n",
+        "                        world_size = dist.get_world_size()\n",
+        "                        rank = dist.get_rank()\n",
         "\n",
-        "                    # Reshape if needed\n",
-        "                    if len(orig_shape) == 4:\n",
-        "                        update = update.view(orig_shape)\n",
+        "                        # Gather update tensor from all ranks\n",
+        "                        update_list = [torch.empty_like(update) for _ in range(world_size)]\n",
+        "                        dist.all_gather(update_list, update)\n",
+        "\n",
+        "                        # Concatenate to get full update\n",
+        "                        full_update = torch.cat(update_list, dim=0)  # Assuming sharding on dim 0\n",
+        "\n",
+        "                        # Perform Newton-Schulz on full matrix\n",
+        "                        orig_shape = full_update.shape\n",
+        "                        if full_update.ndim == 4:\n",
+        "                            full_update = full_update.view(full_update.shape[0], -1)\n",
+        "\n",
+        "                        full_update = self.zeropower_via_newtonschulz5(full_update, steps=ns_steps)\n",
+        "                        full_update *= max(1, grad.size(-2) / grad.size(-1)) ** 0.5\n",
+        "\n",
+        "                        if len(orig_shape) == 4:\n",
+        "                            full_update = full_update.view(orig_shape)\n",
+        "\n",
+        "                        # Extract this rank's shard from the orthogonalized update\n",
+        "                        shard_size = full_update.shape[0] // world_size\n",
+        "                        start_idx = rank * shard_size\n",
+        "                        end_idx = start_idx + shard_size\n",
+        "                        update = full_update[start_idx:end_idx]\n",
+        "\n",
+        "                    else:\n",
+        "                        # Non-FSDP path (single GPU)\n",
+        "                        # Handle conv filters\n",
+        "                        orig_shape = update.shape\n",
+        "                        if update.ndim == 4:\n",
+        "                            update = update.view(update.shape[0], -1)\n",
+        "\n",
+        "                        # Apply Newton-Schulz\n",
+        "                        update = self.zeropower_via_newtonschulz5(update, steps=ns_steps)\n",
+        "\n",
+        "                        # Scale update\n",
+        "                        update *= max(1, grad.size(-2) / grad.size(-1)) ** 0.5\n",
+        "\n",
+        "                        # Reshape if needed\n",
+        "                        if len(orig_shape) == 4:\n",
+        "                            update = update.view(orig_shape)\n",
         "\n",
         "                else:\n",
         "                    # --- AdamW Update Logic ---\n",
@@ -689,7 +726,7 @@
       "metadata": {
         "id": "UgBBhlu8YSOD"
       },
-      "execution_count": 9,
+      "execution_count": 3,
       "outputs": []
     },
     {
@@ -746,16 +783,16 @@
           "base_uri": "https://localhost:8080/"
         },
         "id": "JsLd9EUbYfMw",
-        "outputId": "447510b5-446c-48da-b10f-5ee35d1e137e"
+        "outputId": "cc23c0fd-a6fb-4a38-bd59-96b198bdb3f9"
       },
-      "execution_count": 12,
+      "execution_count": 4,
       "outputs": [
         {
           "output_type": "stream",
           "name": "stdout",
           "text": [
-            "Gradient norm: 40.4564\n",
-            "Weight change: 0.0680\n"
+            "Gradient norm: 37.9263\n",
+            "Weight change: 0.0685\n"
           ]
         }
       ]
@@ -1128,16 +1165,28 @@
           "base_uri": "https://localhost:8080/"
         },
         "id": "CrWv9OuRYfHl",
-        "outputId": "4a2ce32e-d9b8-43f3-ec0d-9c4f10a770ec"
+        "outputId": "c73f5e9f-f4ca-41fa-f8c1-db840f341800"
       },
-      "execution_count": 13,
+      "execution_count": 5,
       "outputs": [
+        {
+          "output_type": "stream",
+          "name": "stdout",
+          "text": [
+            "Running MuonW Optimizer Tests\n",
+            "==================================================\n",
+            "✓ Conv filters handled correctly\n",
+            "✓ Embedding parameters use AdamW update\n",
+            "✓ Matrix parameters use Muon update\n",
+            "✓ Multiple parameter groups work correctly\n"
+          ]
+        },
         {
           "output_type": "stream",
           "name": "stderr",
           "text": [
             "----------------------------------------------------------------------\n",
-            "Ran 8 tests in 0.021s\n",
+            "Ran 8 tests in 0.223s\n",
             "\n",
             "OK\n"
           ]
@@ -1146,12 +1195,6 @@
           "output_type": "stream",
           "name": "stdout",
           "text": [
-            "Running MuonW Optimizer Tests\n",
-            "==================================================\n",
-            "✓ Conv filters handled correctly\n",
-            "✓ Embedding parameters use AdamW update\n",
-            "✓ Matrix parameters use Muon update\n",
-            "✓ Multiple parameter groups work correctly\n",
             "✓ Nesterov momentum works differently from standard momentum\n",
             "✓ Scalar parameters use AdamW update\n",
             "✓ Weight decay applied correctly\n",