Improve Mac/MPS compatibility and device handling

Added dev/runmac_overnight.sh for optimized Mac training. Updated device-specific logic throughout dataloader, GPT, Muon optimizer, and training scripts to avoid CUDA-only features on MPS/CPU (e.g., torch.compile, pin_memory, non_blocking, bfloat16). Relaxed torch version constraints in pyproject.toml and removed Linux/CUDA-specific PyTorch config for better macOS support.

dev/runmac_overnight.sh

@@ -0,0 +1,126 @@
+#!/bin/bash
+# Optimized overnight training for Mac (MPS/Apple Silicon)
+# Expected runtime: 8-12 hours
+# Expected result: Much better chatbot with coherent responses
+
+set -e  # Exit on error
+
+echo "=================================="
+echo "nanochat Mac Overnight Training"
+echo "=================================="
+echo "Started: $(date)"
+echo ""
+
+# Activate virtual environment
+source .venv/bin/activate
+
+# Configuration
+DEPTH=6                    # Bigger model (6 layers vs 4)
+BASE_ITERATIONS=500        # More base training
+MID_ITERATIONS=150         # More midtraining
+SFT_ITERATIONS=150         # More SFT
+DATA_SHARDS=50             # More training data
+
+echo "Configuration:"
+echo "  Model depth: $DEPTH (36.7M → 82M params)"
+echo "  Base iterations: $BASE_ITERATIONS"
+echo "  Mid iterations: $MID_ITERATIONS"
+echo "  SFT iterations: $SFT_ITERATIONS"
+echo "  Data shards: $DATA_SHARDS"
+echo ""
+
+# Clean up old run
+echo "Cleaning up previous training..."
+rm -f report.md
+python -m scripts.report --reset
+
+# Download training data
+echo ""
+echo "Step 1/6: Downloading training data ($DATA_SHARDS shards)..."
+python -m nanochat.dataset -n $DATA_SHARDS
+
+# Download identity conversations
+echo ""
+echo "Step 2/6: Downloading identity conversations..."
+if [ ! -f ~/.cache/nanochat/identity_conversations.jsonl ]; then
+    curl -L -o ~/.cache/nanochat/identity_conversations.jsonl \
+      https://karpathy-public.s3.us-west-2.amazonaws.com/identity_conversations.jsonl
+else
+    echo "  Already downloaded, skipping."
+fi
+
+# Build tokenizer
+echo ""
+echo "Step 3/6: Training tokenizer..."
+python -m nanochat.tokenizer
+
+# Base model training
+echo ""
+echo "Step 4/6: Training base model ($BASE_ITERATIONS iterations)..."
+echo "  This will take ~2-4 hours..."
+python -m scripts.base_train \
+  --depth=$DEPTH \
+  --max_seq_len=1024 \
+  --device_batch_size=1 \
+  --total_batch_size=1024 \
+  --num_iterations=$BASE_ITERATIONS \
+  --eval_every=100 \
+  --eval_tokens=8192 \
+  --core_metric_every=250 \
+  --core_metric_max_per_task=20 \
+  --sample_every=100
+
+# Evaluate base model
+echo ""
+echo "Evaluating base model..."
+python -m scripts.base_loss
+python -m scripts.base_eval
+
+# Midtraining
+echo ""
+echo "Step 5/6: Midtraining ($MID_ITERATIONS iterations)..."
+echo "  This will take ~2-3 hours..."
+python -m scripts.mid_train \
+  --num_iterations=$MID_ITERATIONS \
+  --device_batch_size=1 \
+  --max_seq_len=1024 \
+  --total_batch_size=1024 \
+  --eval_every=50
+
+# SFT training
+echo ""
+echo "Step 6/6: Chat fine-tuning (SFT) ($SFT_ITERATIONS iterations)..."
+echo "  This will take ~2-3 hours..."
+python -m scripts.chat_sft \
+  --num_iterations=$SFT_ITERATIONS \
+  --device_batch_size=1 \
+  --target_examples_per_step=8 \
+  --eval_steps=10
+
+# Final evaluation
+echo ""
+echo "Running final evaluations..."
+python -m scripts.chat_eval -i sft || echo "Chat eval had issues, skipping..."
+
+# Generate report
+echo ""
+echo "Generating final report..."
+python -m scripts.report
+
+# Copy report to current directory
+cp ~/.cache/nanochat/report/report.md ./report_overnight.md
+
+echo ""
+echo "=================================="
+echo "Training Complete!"
+echo "=================================="
+echo "Finished: $(date)"
+echo ""
+echo "Your chatbot is ready! Chat with it:"
+echo "  python -m scripts.chat_cli -i sft"
+echo ""
+echo "Or start the web UI:"
+echo "  python -m scripts.chat_web -i sft"
+echo ""
+echo "Report saved to: report_overnight.md"
+echo "=================================="

nanochat/dataloader.py

@@ -16,6 +16,9 @@ def tokenizing_distributed_data_loader(B, T, split, tokenizer_threads=4, tokeniz
     bos_token = tokenizer.get_bos_token_id()
     # scratch buffer holds the tokens for one iteration
     token_buffer = deque() # we stream tokens on the right and pop from the left
+    # pin_memory and non_blocking only work on CUDA
+    device_type = device if isinstance(device, str) else device.type
+    use_cuda_optimizations = device_type == "cuda"
 
     # infinite iterator over document batches
     def document_batches():
@@ -38,11 +41,11 @@ def tokenizing_distributed_data_loader(B, T, split, tokenizer_threads=4, tokeniz
             batch_index += 1
         # Move tokens from the deque into the scratch buffer
         tokens = [token_buffer.popleft() for _ in range(needed_tokens)]
-        scratch = torch.tensor(tokens, dtype=torch.int64, pin_memory=True)
+        scratch = torch.tensor(tokens, dtype=torch.int64, pin_memory=use_cuda_optimizations)
         # Create the inputs/targets as 1D tensors
         inputs_cpu = scratch[:-1].to(dtype=torch.int32)
         targets_cpu = scratch[1:]
-        # Reshape to 2D and move to GPU async
-        inputs = inputs_cpu.view(B, T).to(device=device, dtype=torch.int32, non_blocking=True)
-        targets = targets_cpu.view(B, T).to(device=device, dtype=torch.int64, non_blocking=True)
+        # Reshape to 2D and move to device (async on CUDA)
+        inputs = inputs_cpu.view(B, T).to(device=device, dtype=torch.int32, non_blocking=use_cuda_optimizations)
+        targets = targets_cpu.view(B, T).to(device=device, dtype=torch.int64, non_blocking=use_cuda_optimizations)
         yield inputs, targets

nanochat/gpt.py

@@ -35,7 +35,13 @@ class GPTConfig:
 
 def norm(x):
     # Purely functional rmsnorm with no learnable params
+    # Fallback for older PyTorch versions that don't have F.rms_norm
+    if hasattr(F, 'rms_norm'):
         return F.rms_norm(x, (x.size(-1),))
+    else:
+        # Manual RMS norm implementation
+        variance = x.pow(2).mean(-1, keepdim=True)
+        return x * torch.rsqrt(variance + 1e-5)
 
 
 def apply_rotary_emb(x, cos, sin):
@@ -211,7 +217,9 @@ class GPT(nn.Module):
         # calculate the rotation frequencies at each (time, channel) pair
         freqs = torch.outer(t, inv_freq)
         cos, sin = freqs.cos(), freqs.sin()
-        cos, sin = cos.bfloat16(), sin.bfloat16() # keep them in bfloat16
+        # Use bfloat16 on CUDA, float32 on MPS/CPU (MPS doesn't support bfloat16)
+        if device.type == "cuda":
+            cos, sin = cos.bfloat16(), sin.bfloat16()
         cos, sin = cos[None, :, None, :], sin[None, :, None, :] # add batch and head dims for later broadcasting
         return cos, sin
 
@@ -244,7 +252,10 @@ class GPT(nn.Module):
             dict(params=embedding_params, lr=embedding_lr * dmodel_lr_scale),
         ]
         adamw_kwargs = dict(betas=(0.8, 0.95), eps=1e-10, weight_decay=weight_decay)
-        AdamWFactory = DistAdamW if ddp else partial(torch.optim.AdamW, fused=True)
+        # fused=True only works on CUDA, not MPS or CPU
+        device_type = self.get_device().type
+        use_fused = device_type == "cuda"
+        AdamWFactory = DistAdamW if ddp else partial(torch.optim.AdamW, fused=use_fused)
         adamw_optimizer = AdamWFactory(adam_groups, **adamw_kwargs)
         # Create the Muon optimizer for the linear layers
         muon_kwargs = dict(lr=matrix_lr, momentum=0.95)
@@ -263,7 +274,9 @@ class GPT(nn.Module):
         # Grab the rotary embeddings for the current sequence length (they are of shape (1, seq_len, 1, head_dim))
         assert T <= self.cos.size(1), f"Sequence length grew beyond the rotary embeddings cache: {T} > {self.cos.size(1)}"
         assert idx.device == self.cos.device, f"Rotary embeddings and idx are on different devices: {idx.device} != {self.cos.device}"
-        assert self.cos.dtype == torch.bfloat16, "Rotary embeddings must be in bfloat16"
+        # Rotary embeddings are bfloat16 on CUDA, float32 on MPS/CPU
+        expected_dtype = torch.bfloat16 if self.cos.device.type == "cuda" else torch.float32
+        assert self.cos.dtype == expected_dtype, f"Rotary embeddings dtype mismatch: expected {expected_dtype}, got {self.cos.dtype}"
         # if kv cache exists, we need to offset the rotary embeddings to the current position in the cache
         T0 = 0 if kv_cache is None else kv_cache.get_pos()
         cos_sin = self.cos[:, T0:T0+T], self.sin[:, T0:T0+T] # truncate cache to current sequence length

nanochat/muon.py

@@ -6,7 +6,6 @@ import torch
 from torch import Tensor
 import torch.distributed as dist
 
-@torch.compile
 def zeropower_via_newtonschulz5(G: Tensor, steps: int) -> Tensor:
     """
     Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
@@ -19,7 +18,11 @@ def zeropower_via_newtonschulz5(G: Tensor, steps: int) -> Tensor:
     """
     assert G.ndim >= 2 # batched Muon implementation by @scottjmaddox, and put into practice in the record by @YouJiacheng
     a, b, c = (3.4445, -4.7750,  2.0315)
+    # Use bfloat16 on CUDA for speed, float32 on MPS/CPU (MPS doesn't support bfloat16)
+    if G.device.type == "cuda":
         X = G.bfloat16()
+    else:
+        X = G.float()
     if G.size(-2) > G.size(-1):
         X = X.mT
 
@@ -35,6 +38,10 @@ def zeropower_via_newtonschulz5(G: Tensor, steps: int) -> Tensor:
         X = X.mT
     return X
 
+# Compile the function for CUDA; use eager mode for MPS/CPU
+# We check at runtime in the optimizer which version to use
+_zeropower_compiled = torch.compile(zeropower_via_newtonschulz5)
+
 class Muon(torch.optim.Optimizer):
     """
     Muon - MomentUm Orthogonalized by Newton-schulz
@@ -65,6 +72,11 @@ class Muon(torch.optim.Optimizer):
             group = dict(params=[p for p in params if p.numel() == size])
             param_groups.append(group)
         super().__init__(param_groups, defaults)
+        # Determine which zeropower function to use based on device
+        # Check the first parameter to determine device
+        first_param = next(iter(params))
+        self._use_compiled = first_param.device.type == "cuda"
+        self._zeropower_fn = _zeropower_compiled if self._use_compiled else zeropower_via_newtonschulz5
 
     @torch.no_grad()
     def step(self):
@@ -79,7 +91,7 @@ class Muon(torch.optim.Optimizer):
                 buf: Tensor = state["momentum_buffer"]
                 buf.lerp_(g, 1 - group["momentum"])
                 g = g.lerp_(buf, group["momentum"]) if group["nesterov"] else buf
-                g = zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
+                g = self._zeropower_fn(g, steps=group["ns_steps"])
                 p.add_(g, alpha=-group["lr"] * max(1, p.size(-2) / p.size(-1))**0.5)
 
 
@@ -122,6 +134,10 @@ class DistMuon(torch.optim.Optimizer):
                 print(f"Muon: Grouping {len(group_params)} params of shape {shape}, device {device}, dtype {dtype}")
             param_groups.append(dict(params=group_params, zero_buffer=torch.zeros_like(group_params[0])))
         super().__init__(param_groups, defaults)
+        # Determine which zeropower function to use based on device
+        first_param = params[0]
+        self._use_compiled = first_param.device.type == "cuda"
+        self._zeropower_fn = _zeropower_compiled if self._use_compiled else zeropower_via_newtonschulz5
 
     @torch.no_grad()
     def step(self):
@@ -173,7 +189,7 @@ class DistMuon(torch.optim.Optimizer):
                     buf: Tensor = state["momentum_buffer"]
                     buf.lerp_(g, 1.0 - group["momentum"])
                     g = g.lerp_(buf, group["momentum"]) if group["nesterov"] else buf
-                    g = zeropower_via_newtonschulz5(g, steps=group["ns_steps"])
+                    g = self._zeropower_fn(g, steps=group["ns_steps"])
                     scale = (max(1.0, p.size(-2) / p.size(-1)) ** 0.5)
                     p.add_(g, alpha=-group["lr"] * scale)
                 # Replicate updated parameters to all ranks

pyproject.toml

@@ -14,7 +14,7 @@ dependencies = [
     "setuptools>=80.9.0",
     "tiktoken>=0.11.0",
     "tokenizers>=0.22.0",
-    "torch>=2.8.0",
+    "torch>=2.0.0,!=2.3.*,!=2.4.*,!=2.5.*,!=2.6.*,!=2.7.*,!=2.8.*,!=2.9.*",
     "uvicorn>=0.36.0",
     "wandb>=0.21.3",
 ]
@@ -44,19 +44,5 @@ python_files = ["test_*.py"]
 python_classes = ["Test*"]
 python_functions = ["test_*"]
 
-# target torch to cuda 12.8
-[tool.uv.sources]
-torch = [
-  { index = "pytorch-cpu", marker = "sys_platform != 'linux'" },
-  { index = "pytorch-cu128", marker = "sys_platform == 'linux'" },
-]
-
-[[tool.uv.index]]
-name = "pytorch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
-explicit = true
-
-[[tool.uv.index]]
-name = "pytorch-cu128"
-url = "https://download.pytorch.org/whl/cu128"
-explicit = true
+# Note: PyTorch configuration removed for macOS compatibility
+# On Linux with CUDA, you may need to reinstall from pytorch.org/whl/cu128

scripts/base_train.py

@@ -109,7 +109,9 @@ with torch.device("meta"):
 model.to_empty(device=device)
 model.init_weights()
 orig_model = model # original, uncompiled model, for saving raw model state_dict
-model = torch.compile(model, dynamic=False) # TODO: dynamic True/False think through
+# torch.compile only works well on CUDA; skip on MPS/CPU
+if device_type == "cuda":
+    model = torch.compile(model, dynamic=False) # TODO: dynamic True/False think through
 num_params = sum(p.numel() for p in model.parameters())
 print0(f"Number of parameters: {num_params:,}")
 num_flops_per_token = model.estimate_flops()

scripts/mid_train.py

@@ -15,6 +15,9 @@ os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
 import time
 import wandb
 import torch
+# Disable torch.compile's dynamo on MPS (not supported)
+import torch._dynamo
+torch._dynamo.config.suppress_errors = True
 from contextlib import nullcontext
 from nanochat.common import compute_init, compute_cleanup, print0, DummyWandb, get_base_dir, autodetect_device_type
 from nanochat.tokenizer import get_token_bytes
@@ -70,7 +73,9 @@ pretrain_batch_size = meta.get("device_batch_size", None)
 if pretrain_batch_size is not None and device_batch_size > pretrain_batch_size:
     print0(f"FOOTGUN WARNING: base model training used device_batch_size {pretrain_batch_size}, did you pass in a good --device_batch_size to this script?")
 orig_model = model
-model = torch.compile(model, dynamic=False)
+# torch.compile only works well on CUDA; skip on MPS/CPU
+if device_type == "cuda":
+    model = torch.compile(model, dynamic=False)
 depth = model.config.n_layer
 num_flops_per_token = model.estimate_flops()
 tokens_per_fwdbwd = device_batch_size * max_seq_len # tokens per iteration for a single rank
@@ -119,7 +124,9 @@ def mid_data_generator(split):
     assert dataset_size > 0
     needed_tokens = device_batch_size * max_seq_len + 1 # to form one training batch of inputs,targets
     token_buffer = deque()
-    scratch = torch.empty(needed_tokens, dtype=torch.int64, pin_memory=True)
+    # pin_memory only works on CUDA
+    use_pin_memory = device_type == "cuda"
+    scratch = torch.empty(needed_tokens, dtype=torch.int64, pin_memory=use_pin_memory)
     cursor = ddp_rank # increments by ddp_world_size each time, so each rank processes unique documents
     it = 0 # iteration counter
     while True:
@@ -142,8 +149,9 @@ def mid_data_generator(split):
             scratch[i] = token_buffer.popleft()
         inputs_cpu = scratch[:-1].to(dtype=torch.int32)
         targets_cpu = scratch[1:]
-        inputs = inputs_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int32, non_blocking=True)
-        targets = targets_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int64, non_blocking=True)
+        # non_blocking only works on CUDA
+        inputs = inputs_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int32, non_blocking=use_pin_memory)
+        targets = targets_cpu.view(device_batch_size, max_seq_len).to(device=device, dtype=torch.int64, non_blocking=use_pin_memory)
         if split == "train":
             if num_iterations > 0:
                 approx_progress = it / num_iterations # calculate progress from the max number of iterations