diff --git a/README.md b/README.md index 08c184a..1894ac8 100644 --- a/README.md +++ b/README.md @@ -3,24 +3,22 @@ ![nanochat logo](dev/nanochat.png) ![scaling laws](dev/scaling_laws_jan26.png) -nanochat is the simplest experimental harness for training LLMs. It is designed to run on a single GPU node, the code is minimal/hackable, and it covers all major LLM stages including tokenization, pretraining, finetuning, evaluation, inference, and a chat UI. For example, you can train your own GPT-2 capability LLM (which cost ~$50,000 to train in 2019) for only $73 (3 hours of 8XH100 GPU node) and then talk to it in a familiar ChatGPT-like web UI. +nanochat is the simplest experimental harness for training LLMs. It is designed to run on a single GPU node, the code is minimal/hackable, and it covers all major LLM stages including tokenization, pretraining, finetuning, evaluation, inference, and a chat UI. For example, you can train your own GPT-2 capability LLM (which cost ~$43,000 to train in 2019) for only $72 (~3 hours of 8XH100 GPU node) and then talk to it in a familiar ChatGPT-like web UI. On a spot instance, the total cost can be closer to ~$20. More generally, nanochat is configured out of the box to train an entire miniseries of compute-optimal models by setting one single complexity dial: `--depth`, the number of layers in the GPT transformer model (GPT-2 capability happens to be approximately depth 26). All other hyperparameters (the width of the transformer, number of heads, learning rate adjustments, training horizons, weight decays, ...) are calculated automatically in an optimal way. For questions about the repo, I recommend either using [DeepWiki](https://deepwiki.com/karpathy/nanochat) from Devin/Cognition to ask questions about the repo, or use the [Discussions tab](https://github.com/karpathy/nanochat/discussions), or come by the [#nanochat](https://discord.com/channels/1020383067459821711/1427295580895314031) channel on Discord. -## Updates +## Time-to-GPT-2 Leaderboard -- (Jan 31 2026) Major revamp of all scripts/README ongoing, deleting midtraining stage, might be a bit messy briefly... -- (Jan 30 2026) With all the latest improvements we're able to train GPT-2 grade LLM in about $73. The [runs/speedrun.sh](runs/speedrun.sh) script will become the refernece way to train GPT-2 grade model and talk to it. - -## Leaderboard +Presently, the main focus of development is on tuning the pretraining stage, which takes the most amount of compute. Inspired by the modded-nanogpt repo and to incentivise progress and community collaboration, nanochat maintains a leaderboard for a "GPT-2 speedrun", which is the wall-clock time required to train a nanochat model to GPT-2 grade capability, as measured by the DCLM CORE score. The [runs/speedrun.sh](runs/speedrun.sh) script always reflects the reference way to train GPT-2 grade model and talk to it. The current leaderboard looks as follows: | # | time | val_bpb | CORE | Description | Date | Commit | Contributors | |---|-------------|---------|------|-------------|------|--------|--------------| | 0 | 168 hours | - | 0.2565 | Original OpenAI GPT-2 checkpoint | 2019 | - | OpenAI | | 1 | 3.04 | 0.74833 | 0.2585 | d24 baseline, slightly overtrained | Jan 29 2026 | 348fbb3 | @karpathy | -| 2 | 2.91 | 0.74504 | 0.2578 | d26 slightly undertrained **+fp8** | Feb 2 2026 | 8309b83 | @karpathy | +| 2 | 2.91 | 0.74504 | 0.2578 | d26 slightly undertrained **+fp8** | Feb 2 2026 | a67eba3 | @karpathy | +| 3 | 2.76 | 0.74645 | 0.2602 | bump total batch size to 1M tokens | Feb 5 2026 | 2c062aa | @karpathy | -The primary metric we care about is "time to GPT-2" - the wall clock time needed to outperform the GPT-2 (1.6B) CORE metric on an 8XH100 GPU node. The GPT-2 CORE score is 0.256525. In 2019, the training of GPT-2 cost approximately $50,000 so it is incredible that due to many advances over 7 years across the stack, we can now do so much faster and for well below $100 (e.g. at the current ~$3/GPU/hr, an 8XH100 node is ~$24/hr, so 3 hours is ~$72). +The primary metric we care about is "time to GPT-2" - the wall clock time needed to outperform the GPT-2 (1.6B) CORE metric on an 8XH100 GPU node. The GPT-2 CORE score is 0.256525. In 2019, the training of GPT-2 cost approximately $43,000 so it is incredible that due to many advances over 7 years across the stack, we can now do so much faster and for well below $100 (e.g. at the current ~$3/GPU/hr, an 8XH100 node is ~$24/hr, so 3 hours is ~$72). See [dev/LEADERBOARD.md](dev/LEADERBOARD.md) for more docs on how to interpret and contribute to the leaderboard. @@ -28,13 +26,13 @@ See [dev/LEADERBOARD.md](dev/LEADERBOARD.md) for more docs on how to interpret a ### Reproduce and talk to GPT-2 -The most fun you can have is to train your own GPT-2 and talk to it. The entire pipeline to do so is contained in the single file [runs/speedrun.sh](runs/speedrun.sh), which is designed to be run on an 8XH100 GPU node. Currently, at ~$24/hour for these nodes, pretraining GPT-2 grade model takes approximately 3 hours and will set you back about $75. Boot up a new 8XH100 GPU box from your favorite provider (e.g. I use and like [Lambda](https://lambda.ai/service/gpu-cloud)), and kick off the training script: +The most fun you can have is to train your own GPT-2 and talk to it. The entire pipeline to do so is contained in the single file [runs/speedrun.sh](runs/speedrun.sh), which is designed to be run on an 8XH100 GPU node. Boot up a new 8XH100 GPU box from your favorite provider (e.g. I use and like [Lambda](https://lambda.ai/service/gpu-cloud)), and kick off the training script: ```bash bash runs/speedrun.sh ``` -You mish to do so in a screen session as this will take ~3 hours to run. Once it's done, you can talk to it via the ChatGPT-like web UI. Make sure again that your local uv virtual environment is active (run `source .venv/bin/activate`), and serve it: +You may wish to do so in a screen session as this will take ~3 hours to run. Once it's done, you can talk to it via the ChatGPT-like web UI. Make sure again that your local uv virtual environment is active (run `source .venv/bin/activate`), and serve it: ```bash python -m scripts.chat_web @@ -69,22 +67,29 @@ OMP_NUM_THREADS=1 torchrun --standalone --nproc_per_node=8 -m scripts.base_train --save-every=-1 \ ``` -This uses wandb (run name "d12"), only runs the CORE metric on last step, and it doesn't sample and save intermediate checkpoints. I like to change something in the code, re-run a d12 (or a d16 etc) and see if it helped, in an iteration loop. +This uses wandb (run name "d12"), only runs the CORE metric on last step, and it doesn't sample and save intermediate checkpoints. I like to change something in the code, re-run a d12 (or a d16 etc) and see if it helped, in an iteration loop. To see if a run helps, I like to monitor the wandb plots for: -The overall approach is to treat the depth of the model as the single dial of complexity. By sweeping out the depth, we get increasingly more powerful models. We determine the scaling laws, set the data budget to a compute optimal setting, train a whole miniseries of models of increasing sizes, and compare them to the GPT-2 and GPT-3 miniseries. Right now, beating GPT-2 specifically faster and faster is the most interesting target. +1. `val_bpb` (validation loss in vocab-size-invariant units of bits per byte) as a function of `step`, `total_training_time` and `total_training_flops`. +2. `core_metric` (the DCLM CORE socre) +3. VRAM utilization, `train/mfu` (Model FLOPS utilization), `train/tok_per_sec` (training throughput) + +See an example [here](https://github.com/karpathy/nanochat/pull/498#issuecomment-3850720044). + +The important thing to note is that nanochat is written and configured around one single dial of complexity - the depth of the transformer. This single integer automatically determines all other hyperparameters (the width of the transformer, number of heads, learning rate adjustments, training horizons, weight decays, ...) so that the trained model comes out compute optimal. The idea is that the user doesn't have to think about or set any of this, they are simply asking for a smaller or bigger model using `--depth`, and everything "just works". By sweeping out the depth, you achieve the nanochat miniseries of compute optimal models at various sizes. GPT-2 capability model (which is of most interest at the moment) happens to be somewhere around d24-d26 range with the current code. But any candidate changes to the repo have to be principled enough that they work for all settings of depth. ## Running on CPU / MPS -The script [runs/runcpu.sh](runs/runcpu.sh) shows a very simple example of running on CPU or Apple Silicon. It dramatically shrinks the LLM tha tis being trained to make things fit into a reasonable time interval of a few ten minutes of training. You will not get strong results in this way. +The script [runs/runcpu.sh](runs/runcpu.sh) shows a very simple example of running on CPU or Apple Silicon. It dramatically shrinks the LLM that is being trained to make things fit into a reasonable time interval of a few ten minutes of training. You will not get strong results in this way. ## Guides -I've published a number of guides that might contain helpful information: +I've published a number of guides that might contain helpful information, most recent to least recent: -- [Oct 13 2025 original nanochat post](https://github.com/karpathy/nanochat/discussions/1) introducing nanochat, though now it contains some deprecated information and the model is a lot older (with worse results) than current master. +- [Feb 1 2026: Beating GPT-2 for <<$100: the nanochat journey](https://github.com/karpathy/nanochat/discussions/481) - [Jan 7 miniseries v1](https://github.com/karpathy/nanochat/discussions/420) documents the first nanochat miniseries of models. -- To customize your nanochat, see [Guide: infusing identity to your nanochat](https://github.com/karpathy/nanochat/discussions/139) in Discussions, which describes how you can tune your nanochat's personality through synthetic data generation and mixing that data into the SFT stage. - To add new abilities to nanochat, see [Guide: counting r in strawberry (and how to add abilities generally)](https://github.com/karpathy/nanochat/discussions/164). +- To customize your nanochat, see [Guide: infusing identity to your nanochat](https://github.com/karpathy/nanochat/discussions/139) in Discussions, which describes how you can tune your nanochat's personality through synthetic data generation and mixing that data into the SFT stage. +- [Oct 13 2025: original nanochat post](https://github.com/karpathy/nanochat/discussions/1) introducing nanochat, though now it contains some deprecated information and the model is a lot older (with worse results) than current master. ## File structure diff --git a/dev/LEADERBOARD.md b/dev/LEADERBOARD.md index 3b61cc6..b8a727f 100644 --- a/dev/LEADERBOARD.md +++ b/dev/LEADERBOARD.md @@ -29,12 +29,12 @@ Note that: - `depth` controls the size of the Transformer - `run` is the wandb name - `model-tag` is the location of the checkpoints on disk -- `device-batch-size` in the ideal world, you want this to be 32 because with sequence length of 2048 (the default) and 8 GPUs we get `32 X 2048 X 8 = 524,288`, which is the total desired batch size determined to work fairly well around this scale. However, for bigger (e.g. d26), 32 is too much and OOMs, so we decrease it by 2 to 16. The `base_train.py` script automatically compensates for this by calculating that it has to use gradient accumulation of 2 to meet the desired total batch size. Therefore, it will fo forward+backward twice and then a single step. Long story short, the ideal value is 32. If that doesn't fit, you decrease it, e.g. 16, 8, etc., keeping it powers of two so that the gradient accumulation math works out neatly. +- `device-batch-size` in the ideal world, you want this to be 32 because with sequence length of 2048 (the default) and 8 GPUs we get `32 X 2048 X 8 = 524,288`, which is the total desired batch size determined to work fairly well around this scale. However, for bigger (e.g. d26), 32 is too much and OOMs, so we decrease it by 2 to 16. The `base_train.py` script automatically compensates for this by calculating that it has to use gradient accumulation of 2 to meet the desired total batch size. Therefore, it will do forward+backward twice and then a single step. Long story short, the ideal value is 32. If that doesn't fit, you decrease it, e.g. 16, 8, etc., keeping it powers of two so that the gradient accumulation math works out neatly. - `sample-every = -1` turns off periodic sampling - `core-metric-max-per-task=-1` means we run the entire CORE eval - `core-metric-every=999999` a bit of a hacky way to make the CORE eval only happen a single time at the very end of the run - `target-param-data-ratio=8.25` controls the training horizon, which is determined in the script by taking the number of non-embedding model parameters and simply multiplying by this number. The current optimal Tokens:Params ratio can be seen in the defaults of the `base_train.py` script (it is 10.5). 10.5 would produce the *compute optimal* model given the currently measured scaling laws. However, GPT-2 capability is currently somewhere in between a d24 and d26. So to reach it exactly, we want to either overtrain d24 or undertrain d26. In this particular example, I am choosing to slightly undertrain a d26. Note that odd depths (e.g. d25) are not super recommended to use because the math around the transformer sizing and its head dimensions doesn't come out neatly. -- `--fp8` turns on fp8 training. If you GPU does not support fp8, you can leave this out and the code will simply train in bf16. bf16 is higher precision than fp8, so you can actually expect that you might be able to do fewer steps (lower the `target-param-data-ratio`) to achieve the same capability. +- `--fp8` turns on fp8 training. If your GPU does not support fp8, you can leave this out and the code will simply train in bf16. bf16 is higher precision than fp8, so you can actually expect that you might be able to do fewer steps (lower the `target-param-data-ratio`) to achieve the same capability. Once you kick off the run, you wait ~3 hours and then at the end you'll see something like: @@ -46,9 +46,9 @@ wandb: total_training_flops 4.330784131228946e+19 wandb: total_training_time 10949.46713 ``` -Your CORE metric must be greater than GPT-2 0.256525. Then you report the `total_training_time`, (e.g. 10949) which is the time of the training iterations alone, excluding all the evaluations and logging, in seconds. So here for example here it is roughly 10949/60/60 ~= 3.04 hours. You should also note and report the validation bpb of your run because the CORE metric can be a little bit noisy. +Your CORE metric must be greater than GPT-2 0.256525. Then you report the `total_training_time`, (e.g. 10949) which is the time of the training iterations alone, excluding all the evaluations and logging, in seconds. So here for example it is roughly 10949/60/60 ~= 3.04 hours. You should also note and report the validation bpb of your run because the CORE metric can be a little bit noisy. -If you outperform GPT-2 and the time is less than current SOTA in the Leaderboard, you get to make a PR. In addition to raw gains, there are some qualitative and aesthetic considerations that go into whether your improvement is merged. For example, if it is gnarly or it significantly bloats the code, or it seems too esoteric, then we will way those things against the improvement demonstrated. Additionally, nanochat cares not only about targeting a single model, but an entire miniseries of models. So your change must be principled enough that it can easily generalize to other model depths, so that we can sweep out a miniseries. +If you outperform GPT-2 and the time is less than current SOTA in the Leaderboard, you get to make a PR. In addition to raw gains, there are some qualitative and aesthetic considerations that go into whether your improvement is merged. For example, if it is gnarly or it significantly bloats the code, or it seems too esoteric, then we will weigh those things against the improvement demonstrated. Additionally, nanochat cares not only about targeting a single model, but an entire miniseries of models. So your change must be principled enough that it can easily generalize to other model depths, so that we can sweep out a miniseries. After you create the commit, to get the current short git commit hash: @@ -89,7 +89,7 @@ Detailed writeup: [Beating GPT-2 for <<$100: the nanochat journey](https://githu ## Run 2 -Achieved Feb 2 2026 on commit `8309b83`. The launch command was +Achieved Feb 2 2026 on commit `a67eba3`. The launch command was ``` OMP_NUM_THREADS=1 torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- \ @@ -117,3 +117,33 @@ Minimum validation bpb: 0.745036 The big change in this run is `--fp8`, which causes all Linear layers (other than the gates) to be switched to fp8 training using `torchao` with tensorwise fp8 scaling. Each step is of slightly lower quality, but we are taking them a lot faster, coming out net ahead. Anyone who does not have fp8 (e.g. using a GPU without it) can simply leave out the `--fp8` flag to train in bfloat16. This will work just fine but it will produce a slightly stronger model than GPT-2 because of the fp8 -> bf16 precision upgrade. It's possible that one can further tune which layers to include in the fp8 conversion and that e.g. some of the smaller matmuls should be just kept in bf16 etc. Previous record was 3.04 hours, so 2.91 hours is `(3.04 - 2.91)/3.04*100` ~= 4.3% speed improvement. + +## Run 3 + +Achieved Feb 5 2026 on commit `2c062aa`. Launch command: + +``` +OMP_NUM_THREADS=1 torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- \ + --depth=26 \ + --run="d26_feb4_double_batch_ratio8.25" \ + --model-tag="d26_feb4_double_batch_ratio8.25" \ + --device-batch-size=16 \ + --total-batch-size=1048576 \ + --sample-every=-1 \ + --save-every=-1 \ + --core-metric-max-per-task=-1 \ + --core-metric-every=999999 \ + --target-param-data-ratio=8.25 \ + --fp8 +``` + +Result: + +``` +core_metric 0.26024 +step 7226 +total_training_time 9922 +Minimum validation bpb: 0.74645 +``` + +The big change here is that the batch size was doubled from 0.5M to 1M, which works better for a d26 model and allowed me to decrease the number of optimization steps a bit via `--target-param-data-ratio` from 8.5 to 8.25. The TLDR is that the original batch size of 0.5M was tuned for d12, but bigger models (e.g. d26) prefer larger total batch size. I determined in experiments that d26 prefers 1M. Then I implemented and merged a principled way to calculate the optimal batch size given depth so that all nanochat models of all depths benefit. See [dev/LOG.md](dev/LOG.md) entry "2026-02-05: Auto Batch Size Scaling" for more detail. diff --git a/dev/LOG.md b/dev/LOG.md index 908fac1..fce90fd 100644 --- a/dev/LOG.md +++ b/dev/LOG.md @@ -4,6 +4,171 @@ A running summary documenting some experiments and findings. Started ~Jan 7 2026 --- +## 2026-02-19: Mixture of Experts (negative) + +Implemented a DeepSeekV3-style Mixture of Experts layer as a drop-in replacement for the dense MLP. The MoE branch works and improves per-step validation loss, but is not a net improvement on wall clock time due to MoE overhead (at least for our scale of interest of approx GPT-2 capability). + +### Implementation + +Follows DeepSeekV3 and using torchtitan as reference: + +- **8 routed experts, top-2 routing** with sigmoid gating (not softmax) +- **1 shared expert** (dense MLP processing all tokens, following DeepSeekV3) +- **Auxiliary-loss-free load balancing** (DeepSeekV3's expert bias nudging) +- **Iso-FLOP sizing**: `expert_hidden_dim = round(4 * dim / (top_k + num_shared) / 128) * 128`, so active FLOPs per token match the dense MLP +- **`torch._grouped_mm`** for dispatching tokens to experts in a single kernel (instead of a Python for-loop) +- **3D expert weight tensors** `(num_experts, hidden, dim)` — Muon's Polar Express operates on the last two dims, so each expert is independently orthogonalized +- **Active parameter counting** for scaling laws (only `top_k + shared` experts, not all 8) + +### What was easy + +- The core MoE forward pass: router, sort tokens by expert, grouped matmul, scatter back. Conceptually clean. +- Shared expert: just an `nn.Linear` MLP that runs on all tokens alongside the routed path. +- 3D expert params + Muon: only required fixing `second_momentum_buffer` shape to preserve leading dims. +- Load balancing: DeepSeekV3's bias nudging is simple and effective (~10 lines). + +### What was hard / ugly + +- **`torch._grouped_mm` quirks**: requires bf16 (not fp32), column-major right operand, int32 cumulative offsets. The API is undocumented and only discoverable by trial and error. +- **Token count padding**: torchtitan pads each expert's token count to alignment multiples (8 for bf16) for better grouped_mm throughput. We implemented this with both a pure PyTorch approach and a copy of torchtitan's Triton kernel. Both compiled cleanly (0 graph breaks), but with ~65K tokens across 8 experts, each expert already gets ~8K tokens which is well-aligned. The padding overhead (gather/scatter) actually regressed MFU from 35% to 33%. Reverted. +- **FP8 + MoE**: `torch._grouped_mm` does NOT support FP8. There's a separate `torch._scaled_grouped_mm` API that requires per-row scaling (not per-tensor like our `Float8Linear`). The backward pass for weight gradients needs per-group column-wise scaling, which torchao implements with custom Triton kernels. We investigated thoroughly (see `dev/moe_fp8.md`) but did not implement — would require either depending on `torchao.prototype` (unstable) or writing ~200 lines of custom autograd + quantization code. Partial FP8 support exists: the shared expert's `nn.Linear` layers do get converted, but the routed experts (3D `nn.Parameter`) stay in bf16. + +### Results + +- d18: MFU dropped from ~46% to ~35% (the grouped_mm dispatch + token sorting overhead is significant) +- Per-step improvement in validation loss does not compensate for the throughput hit +- Net negative on wall clock time + +### What remains (if revisited) + +- **FP8 for routed experts**: Use `torch._scaled_grouped_mm` with a custom `_Float8GroupedMatmul` autograd function, with bf16 fallback for weight gradient (avoiding the per-group column-wise Triton kernels). + +What's really needed is a fused "FlashMoE" kernel that handles routing + expert dispatch + matmul in one shot (like FlashAttention did for attention), with all the needed features. This doesn't exist yet. Rawdogging MoE with current PyTorch primitives is painful — lots of sorting, gathering, scattering, and layout wrangling around the actual compute. + +### Verdict + +MoE is not worth the trouble for nanochat right now. The code bloat is substantial (moe.py, router, shared expert, load balancing, optimizer fixes, FP8 gaps, active param counting) and the performance is worse wall-clock at our scale of interest. The fundamental issue is that the grouped_mm dispatch overhead eats the FLOP savings from sparsity, at least at our model scales and sequence lengths. + +--- + +## 2026-02-17: Pretraining Data: FineWeb (negative) + +Tried vanilla fineweb instead of fineweb-edu dataset. Significantly, shockingly worse results: + +- d26 (GPT-2): CORE 0.2602 → 0.2241 + +This is the fifth failed attempt to beat pure FineWeb-EDU on CORE score. + +--- + +## 2026-02-17: Pretraining Data Mixture Experiment (negative) + +Tried [hynky/finepdfs_50BT-dclm_30BT-fineweb_edu_20BT](https://huggingface.co/datasets/hynky/finepdfs_50BT-dclm_30BT-fineweb_edu_20BT), a mixture of FinePDFs, DCLM, and FineWeb-EDU. Slightly worse on both model sizes tested: + +- d26 (GPT-2): CORE 0.2602 → 0.2549 +- d18: CORE 0.199 → 0.192 + +This is the fourth failed attempt to beat pure FineWeb-EDU on CORE score. + +--- + +## 2026-02-16: SFT Script Upgrades + +Brought `chat_sft.py` up to parity with `base_train.py` and tuned settings based on SFT sweeps. + +Tuning: + +- **Optimizer warm-start** (`--load-optimizer=1`, default on): loads pretrained momentum buffers via new `load_optimizer_state()` in `checkpoint_manager.py`. LRs are reset to fresh SFT values after load. Loading the optimizer works slightly better but not by too much. +- **LR schedule**: replaced "constant 80%, linear to 0" with warmup/constant/warmdown matching `base_train.py` (`--warmup-ratio`, `--warmdown-ratio`, `--init-lr-frac`, `--final-lr-frac`). Similar to pretraining, warmdown ratio of 0.5 worked the best. `--init-lr-frac` changed from 1.0 slightly lower to 0.8. +- **LR tuning**: attempted to tune all the individual LRs (e.g. does SFT prefer lower LR for embeddings? etc.) but all of this produced negative results. +- **Data mixture**: MMLU epochs 1→3, GSM8K epochs 2→4 (confirmed best from sweeps). Epoch counts now configurable via `--mmlu-epochs` / `--gsm8k-epochs`. Might remove these in the future though. + +Quality of life, footguns, minor fixes: + +- **Hyperparameter inheritance**: SFT now inherits batch sizes and LRs from the pretrained checkpoint metadata by default (CLI overrides still work). Also saved `total_batch_size` to `base_train.py` checkpoint metadata. +- **GC management**: disabled Python GC after step 1 to avoid ~500ms pauses (manual collect every 5000 steps), same as base pretraining. +- **ChatCORE eval**: periodic eval during SFT (`--chatcore-every=200`) across all 6 tasks, logged to wandb. +- **MFU**: uses `get_peak_flops()` for actual GPU instead of hardcoded H100 value. +- Removed `--dry-run` and `--dtype` flags. All ranks now participate in checkpoint save. + +--- + +## 2026-02-05: Auto Batch Size Scaling + +### Background + +So far, the `--total-batch-size` was hardcoded to be `2**19 = 524,288` ~= 0.5M tokens. This was the optimal setting for d12, but when I tried to re-tune it for d26 (GPT-2), I noticed that the optimal was closer to `2**20 = 1,048,576` ~= 1M tokens. This is to be expected - larger models prefer a higher optimal total batch size. However, we have to make sure that all settings of `--depth` get their own optimal batch size calculated in some principled way. Here, I referenced the "Power Lines" paper from Cerebras ([arXiv:2505.13738](https://arxiv.org/abs/2505.13738)) for a lot of related experimentation. In particular, they found that **Bopt ∝ D^0.383** (where D is the number of training tokens, not the number of parameters!). So the idea is to tune the optimal batch size on d12, and then extrapolate it with this power law to bigger models. The 0.383 exponent means batch size grows slowly: 10× more tokens only justifies ~2.4× bigger batch. For nanochat's compute-optimal training (D ∝ N via `--target-param-data-ratio`), this means deeper models naturally want larger batches. + +### Implementation + +Added `--total-batch-size=-1` (now the default) to auto-compute optimal batch: + +```python +get_scaling_params = lambda m: m.num_scaling_params()['transformer_matrices'] + m.num_scaling_params()['lm_head'] +if args.total_batch_size == -1: + D_REF = args.target_param_data_ratio * get_scaling_params(build_model_meta(12)) + B_REF = 2**19 + args.total_batch_size = 2 ** round(math.log2(B_REF * (target_tokens / D_REF) ** 0.383)) +``` + +Reference point: d=12 model with B=2^19 (empirically validated). The reference is computed dynamically so that if the architecture changes (e.g., different `--aspect-ratio`), the math automatically adjusts. However, if the model actually does change too much, one would also want to re-tune the optimal batch size for d=12. + +### Results + +With this formula, we currently get: + +| Depth | Scaling Params | Target Tokens | Auto Batch | +|-------|---------------|---------------|------------| +| d=8 | 42M | 0.44B | 2^18 = 262K | +| d=10-16 | 70M-235M | 0.7B-2.5B | 2^19 = 524K | +| d=18-26 | 324M-918M | 3.4B-9.6B | 2^20 = 1.05M | +| d=32-50 | 1.7B-6.2B | 17.6B-65.6B | 2^21 = 2.1M | + +In particular, this matches empirical observations that d26 prefers ~2^20 while d12 prefers ~2^19. + +### Code Cleanup + +Also refactored model initialization to use `build_model_meta(depth)` helper and `dataclasses.asdict()` for cleaner config handling. + +### Useful references + +- [Bergsma et al., Power Laws for Batch Size, Model Size, and Training Horizon](https://arxiv.org/abs/2505.13738) +- [McCandlish et al., An Empirical Model of Large-Batch Training](https://arxiv.org/abs/1812.06162) +- [Brown et al., Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165) +- [Merrill et al., The Batch Size–Critical Batch Size Myth](https://arxiv.org/abs/2505.23971) + +### One more thing (batch size ramp) + +Tried batch size ramping. The simplest implementation I could think of "tricks" the existing training loop by slicing each micro-batch into smaller pieces and calling optimizer.step() more frequently early in training (1/8 → 1/4 → 1/2 → full batch over the first x% of training, with sqrt LR scaling). Also required a torch.compile warmup phase to pre-compile all slice sizes and avoid recompilation spikes during training. While the idea is sound and small gains were observed, they weren't sufficient to justify the code complexity introduced (conditional slicing logic, warmup with state save/restore, etc.). Not merged for now. + +--- + +## 2026-02-05: SwiGLU Activation (Negative Result) + +Replaced ReLU² MLP activation with SwiGLU (inspired by [twitter](https://x.com/_xjdr/status/2019141521690567058)). SwiGLU uses three projections instead of two, so to match parameters and FLOPs we scale hidden_dim from 4× to 8/3×: + +```python +# Old ReLU²: 2 matrices, 4x expansion +# params: 2 × n × 4n = 8n² +# flops: 2 × 2n × 4n = 16n² per token +self.c_fc = Linear(n_embd, 4 * n_embd) +self.c_proj = Linear(4 * n_embd, n_embd) +x = c_proj(relu(c_fc(x)).square()) + +# New SwiGLU: 3 matrices, 8/3x expansion +# params: 2 × n × (8n/3) + (8n/3) × n = 8n² ✓ matches +# flops: 3 × 2n × (8n/3) = 16n² per token ✓ matches +hidden_dim = (8 * n_embd) // 3 +self.w1 = Linear(n_embd, hidden_dim) # gate +self.w2 = Linear(n_embd, hidden_dim) # up +self.w3 = Linear(hidden_dim, n_embd) # down +x = w3(silu(w1(x)) * w2(x)) +``` + +Tested at both d12 and d24 (GPT-2 scale). Worse on all measures — step efficiency, wall clock time, and FLOPs. ReLU² remains superior for nanochat. **Not adopted.** + +--- + ## 2026-02-03: Flip Muon MLP LR Multiplier (PR #492) Tested flipping the shape-based LR heuristic in Muon from boosting tall matrices (input projections like `c_fc`) to boosting wide matrices (output projections like `c_proj`). The original code applies `max(1, rows/cols)^0.5`, giving ~2x LR to `c_fc`. The flipped version gives ~2x LR to `c_proj` instead, which aligns with classical fan-in/fan-out scaling conventions. This was proposed in [PR #492](https://github.com/karpathy/nanochat/pull/492) and showed improvements in modded-nanogpt. @@ -584,7 +749,7 @@ See the branch `fp8_attempt_fail` for: ### Open Questions - Why does the custom op approach use more memory than vanilla BF16? -- Why is the bump in tok_per_sec so low? We should see ~1.6X speedup in both the forward pass and also (twice) in backward pass for the gradients. Granted, Ahmdal's law is part of the solution because our vocab_size is only 32K so the final layer isn't a huge part of the profile but the expected speedup is still not fully realized. +- Why is the bump in tok_per_sec so low? We should see ~1.6X speedup in both the forward pass and also (twice) in backward pass for the gradients. Granted, Amdahl's law is part of the solution because our vocab_size is only 32K so the final layer isn't a huge part of the profile but the expected speedup is still not fully realized. **Conclusion:** Negative result for now. The implementation works correctly but provides marginal speedup with *increased* memory usage. I'm not understanding the torch.compile interaction here. The complexity of FP8 custom ops isn't justified for lm_head alone. TODO to study in more detail the way this is implemented in other libraries, e.g. torchao. @@ -733,8 +898,8 @@ Cherry-picked improvements from NorMuon (modded-nanogpt) into our simpler Muon i - Both methods kept in code for easy comparison (`zeropower_via_polar_express` vs `zeropower_via_newtonschulz5`) - **Result:** No dramatic/noticeable difference in training, but keeping the new Polar Express as default. -**2. Variance Reduction (NorMuon-style)** -- Added low-rank variance estimator similar to Adafactor ([arxiv.org/pdf/2510.05491](https://arxiv.org/pdf/2510.05491)) +**2. NorMuon Variance Reduction** +- Added per-neuron/column adaptive learning rate from NorMuon ([arxiv.org/pdf/2510.05491](https://arxiv.org/pdf/2510.05491)) - Maintains `second_momentum_buffer` with shape `[rows, 1]` or `[1, cols]` (whichever is smaller) - Normalizes updates based on running per-row/col variance estimate (beta2=0.95) - Memory overhead: ~1/max(rows, cols) per param, negligible @@ -748,7 +913,7 @@ Cherry-picked improvements from NorMuon (modded-nanogpt) into our simpler Muon i - Now defaults to ON for Muon via the `weight_decay` param. AdamW still has no weight decay and is hardcoded to 0 weight decay, might try to re-tune this later. **4. Weight decay schedule** -- Added a linear schedule to weight decay that is default on from 1.0 to 0.0 (i.e. start with max weight decay in the beginning of training, them ramp to 0 by the end). Worked better than a static setting in experiments. (modded-nanogpt has the same schedule but it is imlpemented in a more confusing way by multiplying twice by the learning rate, which is already wired up to a decay schedule). +- Added a linear schedule to weight decay that is default on from 1.0 to 0.0 (i.e. start with max weight decay in the beginning of training, then ramp to 0 by the end). Worked better than a static setting in experiments. (modded-nanogpt has the same schedule but it is implemented in a more confusing way by multiplying twice by the learning rate, which is already wired up to a decay schedule). ### Weight Decay Scaling Experiments @@ -776,7 +941,7 @@ Example: If d12 optimal is 0.22, then d20 optimal ≈ 0.22 × (12/20)² ≈ 0.08 ### Summary -Muon was changed to use Polar Express, added Adafactor-style variance reduction, and cautious weight decay with schedule that ramps linearly to zero. All of these changes follow modded-nanogpt repo, but all of them were also validated piece by piece to yield improvements in nanochat with the exception of the Polar Express change which was in the noise. This is default on and configurable with `--weight_decay`, using simply 0.2 and ∝ 1/width² scaling. The kwarg `--weight_decay` is therefore changing as of this change. It used to configure AdamW via standard weight decay and now it becomes exclusively used in Muon (AdamW is hardcoded to 0.0), and it is scaled based on depth. +Muon was changed to use Polar Express, added NorMuon variance reduction, and cautious weight decay with schedule that ramps linearly to zero. All of these changes follow modded-nanogpt repo, but all of them were also validated piece by piece to yield improvements in nanochat with the exception of the Polar Express change which was in the noise. This is default on and configurable with `--weight_decay`, using simply 0.2 and ∝ 1/width² scaling. The kwarg `--weight_decay` is therefore changing as of this change. It used to configure AdamW via standard weight decay and now it becomes exclusively used in Muon (AdamW is hardcoded to 0.0), and it is scaled based on depth. --- @@ -792,6 +957,6 @@ Muon was changed to use Polar Express, added Adafactor-style variance reduction, **Bug Found:** Original implementation clipped local gradients before sync. Since this codebase doesn't use DDP (gradient sync is in the optimizers), each rank was clipping based on its own local norm. Fixed on the branch with proper distributed all-reduce. -**Observartion:** modded-nanogpt does not appear to clip either right now. +**Observation:** modded-nanogpt does not appear to clip either right now. **Summary:** Deleted all grad-clip code paths. The code naturally produces well-behaved gradients. This improves a bit of MFU because we don't have to calculate and sync grad norms. diff --git a/nanochat/checkpoint_manager.py b/nanochat/checkpoint_manager.py index 5a95fbf..f71524e 100644 --- a/nanochat/checkpoint_manager.py +++ b/nanochat/checkpoint_manager.py @@ -170,3 +170,25 @@ def load_model(source, *args, **kwargs): base_dir = get_base_dir() checkpoints_dir = os.path.join(base_dir, model_dir) return load_model_from_dir(checkpoints_dir, *args, **kwargs) + +def load_optimizer_state(source, device, rank, model_tag=None, step=None): + """Load just the optimizer shard for a given rank, without re-loading the model.""" + model_dir = { + "base": "base_checkpoints", + "sft": "chatsft_checkpoints", + "rl": "chatrl_checkpoints", + }[source] + base_dir = get_base_dir() + checkpoints_dir = os.path.join(base_dir, model_dir) + if model_tag is None: + model_tag = find_largest_model(checkpoints_dir) + checkpoint_dir = os.path.join(checkpoints_dir, model_tag) + if step is None: + step = find_last_step(checkpoint_dir) + optimizer_path = os.path.join(checkpoint_dir, f"optim_{step:06d}_rank{rank:d}.pt") + if not os.path.exists(optimizer_path): + log0(f"Optimizer checkpoint not found: {optimizer_path}") + return None + log0(f"Loading optimizer state from {optimizer_path}") + optimizer_data = torch.load(optimizer_path, map_location=device) + return optimizer_data diff --git a/nanochat/common.py b/nanochat/common.py index 9bcd5dd..2dd0792 100644 --- a/nanochat/common.py +++ b/nanochat/common.py @@ -170,7 +170,7 @@ def compute_init(device_type="cuda"): # cuda|cpu|mps # Precision if device_type == "cuda": - torch.backends.fp32_precision = "tf32" # uses tf32 instead of fp32 for matmuls + torch.set_float32_matmul_precision("high") # uses tf32 instead of fp32 for matmuls, see https://docs.pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html # Distributed setup: Distributed Data Parallel (DDP), optional, and requires CUDA is_ddp_requested, ddp_rank, ddp_local_rank, ddp_world_size = get_dist_info() diff --git a/nanochat/fp8.py b/nanochat/fp8.py new file mode 100644 index 0000000..3e88285 --- /dev/null +++ b/nanochat/fp8.py @@ -0,0 +1,266 @@ +"""Minimal FP8 training for nanochat — tensorwise dynamic scaling only. + +Drop-in replacement for torchao's Float8Linear (~2000 lines) with ~150 lines. +We only need the "tensorwise" recipe (one scalar scale per tensor), not the full +generality of torchao (rowwise scaling, FSDP float8 all-gather, DTensor, tensor +subclass dispatch tables, etc.) + +How FP8 training works +====================== +A standard Linear layer does one matmul in forward and two in backward: + forward: output = input @ weight.T + backward: grad_input = grad_output @ weight + grad_weight= grad_output.T @ input + +FP8 training wraps each of these three matmuls with: + 1. Compute scale = FP8_MAX / max(|tensor|) for each operand + 2. Quantize: fp8_tensor = clamp(tensor * scale, -FP8_MAX, FP8_MAX).to(fp8) + 3. Matmul via torch._scaled_mm (cuBLAS FP8 kernel, ~2x faster than bf16) + 4. Dequantize: _scaled_mm handles this internally using the inverse scales + +The key insight: torch._scaled_mm and the float8 dtypes are PyTorch built-ins. +torchao is just orchestration around these primitives. We can call them directly. + +FP8 dtype choice +================ +There are two FP8 formats. We use both, following the standard convention: + - float8_e4m3fn: 4-bit exponent, 3-bit mantissa, range [-448, 448] + Higher precision (more mantissa bits), used for input and weight. + - float8_e5m2: 5-bit exponent, 2-bit mantissa, range [-57344, 57344] + Wider range (more exponent bits), used for gradients which can be large. + +torch._scaled_mm layout requirements +===================================== +The cuBLAS FP8 kernel requires specific memory layouts: + - First argument (A): must be row-major (contiguous) + - Second argument (B): must be column-major (B.t().contiguous().t()) +If B is obtained by transposing a contiguous tensor (e.g. weight.t()), it is +already column-major — no copy needed. Otherwise we use _to_col_major(). + +How this differs from torchao's approach +======================================== +torchao uses a "tensor subclass" architecture: Float8TrainingTensor is a subclass +of torch.Tensor that bundles FP8 data + scale + metadata. It implements +__torch_dispatch__ with a dispatch table that intercepts every aten op (mm, t, +reshape, clone, ...) and handles it in FP8-aware fashion. When you call + output = input @ weight.T +the @ operator dispatches to aten.mm, which gets intercepted and routed to +torch._scaled_mm behind the scenes. This is ~2000 lines of code because you need +a handler for every tensor operation that might touch an FP8 tensor. + +We take a simpler approach: a single autograd.Function (_Float8Matmul) that takes +full-precision inputs, quantizes to FP8 internally, calls _scaled_mm, and returns +full-precision outputs. Marked @allow_in_graph so torch.compile treats it as one +opaque node rather than trying to trace inside. + +The trade-off is in how torch.compile sees the two approaches: + - torchao: compile decomposes the tensor subclass (via __tensor_flatten__) and + sees every individual op (amax, scale, cast, _scaled_mm) as separate graph + nodes. Inductor can fuse these with surrounding operations (e.g. fuse the + amax computation with the preceding layer's activation function). + - ours: compile sees a single opaque call. It can optimize everything around + the FP8 linear (attention, norms, etc.) but cannot fuse across the boundary. + +Both call the exact same cuBLAS _scaled_mm kernel — the GPU matmul is identical. +The difference is only in the "glue" ops (amax, scale, cast) which are tiny +compared to the matmul. In practice this means our version is slightly faster +(less compilation overhead, no tensor subclass dispatch cost) but can produce +subtly different floating-point rounding paths under torch.compile, since Inductor +generates a different graph. Numerics are bitwise identical in eager mode. +""" + +import torch +import torch.nn as nn + +# Avoid division by zero when computing scale from an all-zeros tensor +EPS = 1e-12 + + +@torch.no_grad() +def _to_fp8(x, fp8_dtype): + """Dynamically quantize a tensor to FP8 using tensorwise scaling. + + "Tensorwise" means one scalar scale for the entire tensor (as opposed to + "rowwise" which computes a separate scale per row). Tensorwise is faster + because cuBLAS handles the scaling; rowwise needs the CUTLASS kernel. + + Returns (fp8_data, inverse_scale) for use with torch._scaled_mm. + """ + fp8_max = torch.finfo(fp8_dtype).max + # Compute the max absolute value across the entire tensor + amax = x.float().abs().max() + # Scale maps [0, amax] -> [0, fp8_max]. Use float64 for the division to + # ensure consistent numerics between torch.compile and eager mode. + # (torchao does the same upcast — without it, compile/eager can diverge) + scale = fp8_max / amax.double().clamp(min=EPS) + scale = scale.float() + # Quantize: scale into FP8 range, saturate (clamp prevents overflow when + # casting — PyTorch's default is to wrap, not saturate), then cast to FP8 + x_scaled = x.float() * scale + x_clamped = x_scaled.clamp(-fp8_max, fp8_max) + x_fp8 = x_clamped.to(fp8_dtype) + # _scaled_mm expects the *inverse* of our scale (it multiplies by this to + # convert FP8 values back to the original range during the matmul) + inv_scale = scale.reciprocal() + return x_fp8, inv_scale + + +def _to_col_major(x): + """Rearrange a 2D tensor's memory to column-major layout. + + torch._scaled_mm requires its second operand in column-major layout. + The trick: transpose -> contiguous (forces a copy in transposed order) + -> transpose back. The result has the same logical shape but column-major + strides, e.g. a [M, N] tensor gets strides (1, M) instead of (N, 1). + """ + return x.t().contiguous().t() + + +# allow_in_graph tells torch.compile to treat this as an opaque operation — +# dynamo won't try to decompose it into smaller ops. See the module docstring +# for how this differs from torchao's tensor subclass approach. +@torch._dynamo.allow_in_graph +class _Float8Matmul(torch.autograd.Function): + """Custom autograd for the three FP8 GEMMs of a Linear layer. + + The forward quantizes input and weight to FP8 and saves + the quantized tensors + scales for backward. + """ + + @staticmethod + def forward(ctx, input_2d, weight): + # Quantize both operands to e4m3 (higher precision format) + input_fp8, input_inv = _to_fp8(input_2d, torch.float8_e4m3fn) + weight_fp8, weight_inv = _to_fp8(weight, torch.float8_e4m3fn) + ctx.save_for_backward(input_fp8, input_inv, weight_fp8, weight_inv) + + # output = input @ weight.T + # input_fp8 is [B, K] contiguous = row-major (good for first arg) + # weight_fp8 is [N, K] contiguous, so weight_fp8.t() is [K, N] with + # strides (1, K) = column-major (good for second arg, no copy needed!) + output = torch._scaled_mm( + input_fp8, + weight_fp8.t(), + scale_a=input_inv, + scale_b=weight_inv, + out_dtype=input_2d.dtype, + # use_fast_accum=True accumulates the dot products in lower precision. + # Slightly less accurate but measurably faster. Standard practice for + # the forward pass; we use False in backward for more precise gradients. + use_fast_accum=True, + ) + return output + + @staticmethod + def backward(ctx, grad_output): + in_fp8, in_inv, w_fp8, w_inv = ctx.saved_tensors + + # === GEMM 1: grad_input = grad_output @ weight === + # Shapes: [B, N] @ [N, K] -> [B, K] + # Gradients use e5m2 (wider range), weights use e4m3 (higher precision) + go_fp8, go_inv = _to_fp8(grad_output, torch.float8_e5m2) + # go_fp8 is [B, N] contiguous = row-major, good for first arg + # w_fp8 is [N, K] contiguous = row-major, need column-major for second arg + w_col = _to_col_major(w_fp8) + grad_input = torch._scaled_mm( + go_fp8, + w_col, + scale_a=go_inv, + scale_b=w_inv, + out_dtype=grad_output.dtype, + use_fast_accum=False, + ) + + # === GEMM 2: grad_weight = grad_output.T @ input === + # Shapes: [N, B] @ [B, K] -> [N, K] + # go_fp8 is [B, N] contiguous, we need go.T = [N, B] as first arg. + # Transposing gives column-major, but first arg needs row-major, + # so we must call .contiguous() to physically rearrange the memory. + go_T = go_fp8.t().contiguous() # [N, B] row-major + in_col = _to_col_major(in_fp8) # [B, K] column-major + grad_weight = torch._scaled_mm( + go_T, + in_col, + scale_a=go_inv, + scale_b=in_inv, + out_dtype=grad_output.dtype, + use_fast_accum=False, + ) + + return grad_input, grad_weight + + +class Float8Linear(nn.Linear): + """Drop-in nn.Linear replacement that does FP8 compute. + + Weights and biases remain in their original precision (e.g. fp32/bf16). + Only the matmul is performed in FP8 via the _Float8Matmul autograd function. + """ + + def forward(self, input): + # Replicate the autocast behavior of F.linear — when autocast is active, + # we need to manually cast input to the autocast dtype (e.g. bf16), + # since we bypass F.linear's built-in autocast handling. + if torch.is_autocast_enabled(): + input = input.to(torch.get_autocast_gpu_dtype()) + # _scaled_mm only works on 2D tensors, so flatten batch dimensions + orig_shape = input.shape + input_2d = input.reshape(-1, orig_shape[-1]) + output = _Float8Matmul.apply(input_2d, self.weight) + output = output.reshape(*orig_shape[:-1], output.shape[-1]) + if self.bias is not None: + output = output + self.bias.to(output.dtype) + return output + + @classmethod + def from_float(cls, mod): + """Create Float8Linear from nn.Linear, sharing the same weight and bias. + + Uses meta device to avoid allocating a temporary weight tensor — we + create the module shell on meta (shapes/dtypes only, no memory), then + point .weight and .bias to the original module's parameters. + """ + with torch.device("meta"): + new_mod = cls(mod.in_features, mod.out_features, bias=False) + new_mod.weight = mod.weight + new_mod.bias = mod.bias + return new_mod + + +class Float8LinearConfig: + """Minimal config matching torchao's API. Only tensorwise recipe is supported.""" + + @staticmethod + def from_recipe_name(recipe_name): + if recipe_name != "tensorwise": + raise ValueError( + f"Only 'tensorwise' recipe is supported, got '{recipe_name}'. " + f"Rowwise/axiswise recipes require the full torchao library." + ) + return Float8LinearConfig() + + +def convert_to_float8_training(module, *, config=None, module_filter_fn=None): + """Replace nn.Linear layers with Float8Linear throughout a module. + + Walks the module tree in post-order (children before parents) and swaps + each nn.Linear that passes the optional filter. The new Float8Linear shares + the original weight and bias tensors — no copies, no extra memory. + + Args: + module: Root module to convert. + config: Float8LinearConfig (accepted for API compat, only tensorwise supported). + module_filter_fn: Optional filter(module, fqn) -> bool. Only matching Linears + are converted. Common use: skip layers with dims not divisible by 16 + (hardware requirement for FP8 matmuls on H100). + """ + def _convert(mod, prefix=""): + for name, child in mod.named_children(): + fqn = f"{prefix}.{name}" if prefix else name + _convert(child, fqn) + if isinstance(child, nn.Linear) and not isinstance(child, Float8Linear): + if module_filter_fn is None or module_filter_fn(child, fqn): + setattr(mod, name, Float8Linear.from_float(child)) + + _convert(module) + return module diff --git a/nanochat/optim.py b/nanochat/optim.py index 190a1ed..42d862b 100644 --- a/nanochat/optim.py +++ b/nanochat/optim.py @@ -67,6 +67,10 @@ Polar Express Sign Method for orthogonalization. https://arxiv.org/pdf/2505.16932 by Noah Amsel, David Persson, Christopher Musco, Robert M. Gower. +NorMuon variance reduction: per-neuron/column adaptive learning rate that normalizes +update scales after orthogonalization (Muon's output has non-uniform scales across neurons). +https://arxiv.org/pdf/2510.05491 + Some of the changes in nanochat implementation: - Uses a simpler, more general approach to parameter grouping and stacking - Uses a single fused kernel for the momentum -> polar_express -> variance_reduction -> update step @@ -373,6 +377,7 @@ class DistMuonAdamW(torch.optim.Optimizer): param_infos[p] = dict(future=future, grad_slice=grad, is_small=True) else: # Large params: reduce_scatter + assert grad.shape[0] % world_size == 0, f"AdamW reduce_scatter requires shape[0] ({grad.shape[0]}) divisible by world_size ({world_size})" rank_size = grad.shape[0] // world_size grad_slice = torch.empty_like(grad[:rank_size]) future = dist.reduce_scatter_tensor(grad_slice, grad, op=dist.ReduceOp.AVG, async_op=True).get_future() diff --git a/pyproject.toml b/pyproject.toml index bcb674d..8b6fd95 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -20,7 +20,6 @@ dependencies = [ "tiktoken>=0.11.0", "tokenizers>=0.22.0", "torch==2.9.1", - "torchao==0.15.0", "transformers>=4.57.3", "uvicorn>=0.36.0", "wandb>=0.21.3", diff --git a/runs/miniseries.sh b/runs/miniseries.sh index c42544e..01c4459 100644 --- a/runs/miniseries.sh +++ b/runs/miniseries.sh @@ -28,7 +28,7 @@ fi # Series name: from arg, env var, or default to today's date (e.g., jan11) SERIES_NAME="${1:-${SERIES_NAME:-$(date +%b%d | tr '[:upper:]' '[:lower:]')}}" # Depths to train (the "miniseries") -DEPTHS=(10 11 12 13 14 15 16 17 18 19 20) +DEPTHS=(12 14 16 18 20 22 24 26) # Hardware NPROC_PER_NODE="${NPROC_PER_NODE:-8}" # Logging @@ -57,8 +57,15 @@ for d in "${DEPTHS[@]}"; do TAG="${SERIES_NAME}_miniseries_d${d}" START_TIME=$(date +%s) - # Train the model with natural horizon (target_param_data_ratio default) - # No --target-flops, let it use the default ratio from base_train + # Reduce --device-batch-size to avoid OOM at larger depths + if [ $d -ge 28 ]; then + DEVICE_BATCH_SIZE_ARG="--device-batch-size=8" + elif [ $d -ge 20 ]; then + DEVICE_BATCH_SIZE_ARG="--device-batch-size=16" + else + DEVICE_BATCH_SIZE_ARG="--device-batch-size=32" + fi + torchrun --standalone --nproc_per_node=$NPROC_PER_NODE -m scripts.base_train -- \ --depth=$d \ --run="${WANDB_RUN}_d${d}" \ @@ -67,6 +74,7 @@ for d in "${DEPTHS[@]}"; do --core-metric-max-per-task=-1 \ --sample-every=-1 \ --save-every=-1 \ + $DEVICE_BATCH_SIZE_ARG \ 2>&1 | tee "$RESULTS_DIR/${TAG}_train.log" END_TIME=$(date +%s) diff --git a/runs/speedrun.sh b/runs/speedrun.sh index c423ba6..c757253 100644 --- a/runs/speedrun.sh +++ b/runs/speedrun.sh @@ -69,8 +69,8 @@ python -m scripts.tok_eval echo "Waiting for dataset download to complete..." wait $DATASET_DOWNLOAD_PID -# d24 model (slightly overtrained is enough to beat GPT-2 => increase data:params ratio from compute optimal 10.5 (default) to 12) -torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=26 --target-param-data-ratio=8.5 --device-batch-size=16 --fp8 --run=$WANDB_RUN +# d26 model (slightly undertrained to beat GPT-2 => decrease data:params ratio from compute optimal 10.5 (default) to 8.25) +torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=26 --target-param-data-ratio=8.25 --device-batch-size=16 --fp8 --run=$WANDB_RUN # evaluate the model: CORE metric, BPB on train/val, and draw samples torchrun --standalone --nproc_per_node=8 -m scripts.base_eval -- --device-batch-size=16 diff --git a/scripts/base_train.py b/scripts/base_train.py index fa05b60..24091b6 100644 --- a/scripts/base_train.py +++ b/scripts/base_train.py @@ -11,11 +11,14 @@ If you are only on CPU/Macbook, you'll want to train a much much smaller LLM. Ex python -m scripts.base_train --depth=4 --max-seq-len=512 --device-batch-size=1 --eval-tokens=512 --core-metric-every=-1 --total-batch-size=512 --num-iterations=20 """ -import gc import os os.environ["PYTORCH_ALLOC_CONF"] = "expandable_segments:True" -import argparse +import gc +import json import time +import math +import argparse +from dataclasses import asdict from contextlib import nullcontext, contextmanager import wandb @@ -53,8 +56,8 @@ parser.add_argument("--num-iterations", type=int, default=-1, help="explicit num parser.add_argument("--target-flops", type=float, default=-1.0, help="calculate num_iterations to reach target_flops (-1 = disable)") parser.add_argument("--target-param-data-ratio", type=float, default=10.5, help="calculate num_iterations to maintain data:param ratio (Chinchilla=20, -1 = disable)") # Optimization -parser.add_argument("--device-batch-size", type=int, default=32, help="per-device batch size") -parser.add_argument("--total-batch-size", type=int, default=524288, help="total batch size in tokens") +parser.add_argument("--device-batch-size", type=int, default=32, help="per-device batch size. good number to reduce to 16,8,4,... if you OOM on VRAM.") +parser.add_argument("--total-batch-size", type=int, default=-1, help="total batch size in tokens. decent numbers are e.g. 524288. (-1 = auto-compute optimal)") parser.add_argument("--embedding-lr", type=float, default=0.3, help="learning rate for embedding parameters (Adam)") parser.add_argument("--unembedding-lr", type=float, default=0.004, help="learning rate for unembedding parameters (Adam)") parser.add_argument("--weight-decay", type=float, default=0.2, help="cautious weight decay for the Muon optimizer (for weights)") @@ -78,8 +81,8 @@ parser.add_argument("--model-tag", type=str, default=None, help="override model args = parser.parse_args() user_config = vars(args).copy() # for logging # ----------------------------------------------------------------------------- +# Compute init and wandb logging -# Compute init device_type = autodetect_device_type() if args.device_type == "" else args.device_type ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type) master_process = ddp_rank == 0 # this process will do logging, checkpointing etc. @@ -109,65 +112,39 @@ else: print0("WARNING: Recommend using --window-pattern L for full context attention without alternating sliding window patterns.") print0("!" * 80) -# Tokenizer will be useful for evaluation, also we need the vocab size +# ----------------------------------------------------------------------------- +# Tokenizer will be useful for evaluation and also we need the vocab size to init the model tokenizer = get_tokenizer() token_bytes = get_token_bytes(device=device) vocab_size = tokenizer.get_vocab_size() print0(f"Vocab size: {vocab_size:,}") -# Model kwargs are derived from the desired depth of the model -# We nudge model_dim up to the nearest multiple of head_dim to ensure clean division -# (FA3 requires head_dim divisible by 8, and this guarantees head_dim == args.head_dim exactly) -# (For very small depths, this gives a slight "unfair" advantage to models with odd depths) -num_layers = args.depth -base_dim = args.depth * args.aspect_ratio -model_dim = ((base_dim + args.head_dim - 1) // args.head_dim) * args.head_dim -num_heads = model_dim // args.head_dim -num_kv_heads = num_heads # default is 1:1 GQA (Group Query Attention) ratio (i.e. GQA is disabled) -head_dim = model_dim // num_heads -print0(f"num_layers: {num_layers}") -print0(f"model_dim: {model_dim} (base: {base_dim}, nudge: {model_dim - base_dim:+d})") -print0(f"num_heads: {num_heads}") -print0(f"head_dim: {head_dim}") -print0(f"num_kv_heads: {num_kv_heads}") - -# Optimizer / data / training length related hyperparameters -# figure out the needed gradient accumulation to reach the desired total batch size -tokens_per_fwdbwd = args.device_batch_size * args.max_seq_len # tokens per iteration for a single rank -world_tokens_per_fwdbwd = tokens_per_fwdbwd * ddp_world_size # total tokens per iteration for all ranks -assert args.total_batch_size % world_tokens_per_fwdbwd == 0 -grad_accum_steps = args.total_batch_size // world_tokens_per_fwdbwd -print0(f"Tokens / micro-batch / rank: {args.device_batch_size} x {args.max_seq_len} = {tokens_per_fwdbwd:,}") -print0(f"Tokens / micro-batch: {world_tokens_per_fwdbwd:,}") -print0(f"Total batch size {args.total_batch_size:,} => gradient accumulation steps: {grad_accum_steps}") - -# Batch size scaling for learning rates (hyperparameters were tuned at reference batch size 2^19) -batch_lr_scale = 1.0 -reference_batch_size = 2**19 -batch_ratio = args.total_batch_size / reference_batch_size -if batch_ratio != 1.0: - # SGD: linear scaling with batch size is standard (not used in nanochat) - # AdamW: sqrt scaling is standard - # Muon: sqrt scaling is an assumption - not fully studied, but it's a second-order-ish optimizer - batch_lr_scale = batch_ratio ** 0.5 - print0(f"Scaling LRs by {batch_lr_scale:.4f} for batch size {args.total_batch_size:,} (reference: {reference_batch_size:,})") - -# Weight decay is tuned at d12 and its scaling seems to be \propto 1/channels^2 (or equivalently, \propto 1/depth^2 due to constant aspect ratio) -weight_decay_scaled = args.weight_decay * (12 / args.depth)**2 -if args.depth != 12: - print0(f"Scaling weight decay from {args.weight_decay:.6f} to {weight_decay_scaled:.6f} for depth {args.depth}") - # ----------------------------------------------------------------------------- # Initialize the Model -# Create a new model with random weights -model_config_kwargs = dict(sequence_len=args.max_seq_len, vocab_size=vocab_size, n_layer=num_layers, n_head=num_heads, n_kv_head=num_kv_heads, n_embd=model_dim, window_pattern=args.window_pattern) -with torch.device("meta"): - # All tensors are created as meta tensors (they have shape/dtype but no data) - model_config = GPTConfig(**model_config_kwargs) - model = GPT(model_config) -model.to_empty(device=device) # All tensors get storage on target device but with uninitialized (garbage) data -model.init_weights() # All tensors get initialized +def build_model_meta(depth): + """Build a model on meta device for a given depth (shapes/dtypes only, no data).""" + # Model dim is nudged up to nearest multiple of head_dim for clean division + # (FA3 requires head_dim divisible by 8, and this guarantees head_dim == args.head_dim exactly) + base_dim = depth * args.aspect_ratio + model_dim = ((base_dim + args.head_dim - 1) // args.head_dim) * args.head_dim + num_heads = model_dim // args.head_dim + config = GPTConfig( + sequence_len=args.max_seq_len, vocab_size=vocab_size, + n_layer=depth, n_head=num_heads, n_kv_head=num_heads, n_embd=model_dim, + window_pattern=args.window_pattern, + ) + with torch.device("meta"): + model_meta = GPT(config) + return model_meta + +# Build the model, move to device, init the weights +model = build_model_meta(args.depth) # 1) Build on meta device (only shapes/dtypes, no data) +model_config = model.config +model_config_kwargs = asdict(model_config) +print0(f"Model config:\n{json.dumps(model_config_kwargs, indent=2)}") +model.to_empty(device=device) # 2) All tensors get storage on target device but with uninitialized (garbage) data +model.init_weights() # 3) All tensors get initialized # If we are resuming, overwrite the model parameters with those of the checkpoint base_dir = get_base_dir() @@ -181,64 +158,34 @@ if resuming: del model_data # free up this memory after the copy # ----------------------------------------------------------------------------- -# Determine the length of the training run based on model size - -# Detailed parameter counts -param_counts = model.num_scaling_params() -print0(f"Parameter counts:") -for key, value in param_counts.items(): - print0(f"{key:24s}: {value:,}") -num_params = param_counts['total'] -num_scaling_params = param_counts['transformer_matrices'] + param_counts['lm_head'] # determined to give the cleanest scaling laws, see dev/LOG.md Jan 27, 2026 -num_flops_per_token = model.estimate_flops() -print0(f"Estimated FLOPs per token: {num_flops_per_token:e}") - -# Calculate number of iterations. Either it is given, or from target flops, or from target data:param ratio (in that order) -assert args.num_iterations > 0 or args.target_param_data_ratio > 0 or args.target_flops > 0 -if args.num_iterations > 0: - num_iterations = args.num_iterations - print0(f"Using user-provided number of iterations: {num_iterations:,}") -elif args.target_flops > 0: - # calculate the number of iterations from the target flops - num_iterations = round(args.target_flops / (num_flops_per_token * args.total_batch_size)) - print0(f"Calculated number of iterations from target FLOPs: {num_iterations:,}") -elif args.target_param_data_ratio > 0: - # calculate the number of iterations from the target param data ratio (use scaling params per Kaplan et al.) - target_tokens = int(args.target_param_data_ratio * num_scaling_params) - num_iterations = target_tokens // args.total_batch_size - print0(f"Calculated number of iterations from target data:param ratio: {num_iterations:,}") -else: - raise ValueError("No training horizon specified") -total_tokens = args.total_batch_size * num_iterations -print0(f"Total number of training tokens: {total_tokens:,}") -print0(f"Tokens : Scaling params ratio: {args.total_batch_size * num_iterations / num_scaling_params:.2f}") # Chinchilla is ~20 -print0(f"Total training FLOPs estimate: {num_flops_per_token * total_tokens:e}") - -# ----------------------------------------------------------------------------- -# FP8 training initialization and management (has to be done before torch.compile) +# FP8 training initialization and management (this has to be done before torch.compile) # Convert Linear layers to Float8Linear if --fp8 is set if args.fp8: if device_type != "cuda": print0("Warning: FP8 training requires CUDA, ignoring --fp8 flag") else: - from torchao.float8 import Float8LinearConfig, convert_to_float8_training + # our custom fp8 is simpler than torchao, written for exact API compatibility + from nanochat.fp8 import Float8LinearConfig, convert_to_float8_training + # from torchao.float8 import Float8LinearConfig, convert_to_float8_training import torch.nn as nn - # Filter: only convert layers with dimensions divisible by 16 (FP8 hardware requirement) + # Filter: dims must be divisible by 16 (FP8 hardware requirement) large enough def fp8_module_filter(mod: nn.Module, fqn: str) -> bool: if not isinstance(mod, nn.Linear): return False - # FP8 requires both in_features and out_features divisible by 16 if mod.in_features % 16 != 0 or mod.out_features % 16 != 0: return False + if min(mod.in_features, mod.out_features) < 128: + return False return True fp8_config = Float8LinearConfig.from_recipe_name(args.fp8_recipe) + num_linear = sum(1 for m in model.modules() if isinstance(m, nn.Linear)) convert_to_float8_training(model, config=fp8_config, module_filter_fn=fp8_module_filter) - num_fp8_layers = sum(1 for m in model.modules() if 'Float8' in type(m).__name__) - num_skipped = sum(1 for m in model.modules() if isinstance(m, nn.Linear)) - num_fp8_layers - print0(f"✓ FP8 training enabled ({args.fp8_recipe} scaling) - converted {num_fp8_layers} layers, skipped {num_skipped} (dims not divisible by 16)") + num_fp8 = sum(1 for m in model.modules() if 'Float8' in type(m).__name__) + num_skipped = num_linear - num_fp8 + print0(f"✓ FP8 training enabled ({args.fp8_recipe} scaling) - converted {num_fp8}/{num_linear} linear layers, skipped {num_skipped} (too small)") # Context manager to temporarily disable FP8 so that model evaluation remains in BF16 @contextmanager @@ -293,16 +240,75 @@ def disable_fp8(model): orig_model = model # original, uncompiled model, for saving raw model state_dict and for inference/evaluation (because the shapes may change shape) model = torch.compile(model, dynamic=False) # the inputs to model will never change shape so dynamic=False is safe +# ----------------------------------------------------------------------------- +# Scaling laws and muP extrapolations to determine the optimal training horizon, batch size, learning rates, weight decay. + +# Get the parameter counts of our model +param_counts = model.num_scaling_params() +print0(f"Parameter counts:") +for key, value in param_counts.items(): + print0(f"{key:24s}: {value:,}") +num_params = param_counts['total'] +num_flops_per_token = model.estimate_flops() +print0(f"Estimated FLOPs per token: {num_flops_per_token:e}") + +# 1) Use scaling laws to determine the optimal training horizon in tokens +# The compute-optimal models satisfy the Tokens:Params ratio of --target-param-data-ratio (derived experimentally via scaling laws analysis). +# We've already initialized the model so we have Params. Optimal Tokens is now simply target-param-data-ratio * Params +def get_scaling_params(m): + # As for which params to use exactly, transformer matrices + lm_head gives cleanest scaling laws (see dev/LOG.md Jan 27, 2026) + params_counts = m.num_scaling_params() + scaling_params = params_counts['transformer_matrices'] + params_counts['lm_head'] + return scaling_params +num_scaling_params = get_scaling_params(model) +target_tokens = int(args.target_param_data_ratio * num_scaling_params) # optimal tokens for the model we are about to train + +# Our reference model is d12, this is where a lot of hyperparameters are tuned and then transfered to higher depths (muP style) +d12_ref = build_model_meta(12) # creates the model on meta device +D_REF = args.target_param_data_ratio * get_scaling_params(d12_ref) # compute-optimal d12 training horizon in tokens (measured empirically) +B_REF = 2**19 # optimal batch size at d12 ~= 524,288 tokens (measured empirically) + +# 2) Now that we have the token horizon, we can calculate the optimal batch size +# We follow the Power Lines paper (Bopt ∝ D^0.383), ref: https://arxiv.org/abs/2505.13738 +# The optimal batch size grows as approximately D^0.383, so e.g. if D doubles from d12 to d24, B should grow by 2^0.383 ≈ 1.3x. +total_batch_size = args.total_batch_size # user-provided override is possible +if total_batch_size == -1: + batch_size_ratio = target_tokens / D_REF + predicted_batch_size = B_REF * batch_size_ratio ** 0.383 + total_batch_size = 2 ** round(math.log2(predicted_batch_size)) # clamp to nearest power of 2 for efficiency + print0(f"Auto-computed optimal batch size: {total_batch_size:,} tokens") + +# 3) Knowing the batch size, we can now calculate a learning rate correction (bigger batch size allows higher learning rates) +batch_lr_scale = 1.0 +batch_ratio = total_batch_size / B_REF # B/B_ref +if batch_ratio != 1.0: + # SGD: linear scaling with batch size is standard (not used in nanochat) + # AdamW: sqrt scaling is standard: η ∝ √(B/B_ref) + # Muon: we will use the same scaling for Muon as for AdamW: η ∝ √(B/B_ref) (not studied carefully, assumption!) + batch_lr_scale = batch_ratio ** 0.5 # η ∝ √(B/B_ref) + print0(f"Scaling LRs by {batch_lr_scale:.4f} for batch size {total_batch_size:,} (reference: {B_REF:,})") + +# 4) Knowing the batch size and the token horizon, we can now calculate the appropriate weight decay scaling +# We adopt the T_epoch framework from https://arxiv.org/abs/2405.13698 +# Central idea of the paper is that T_epoch = B/(η·λ·D) should remain constant. +# Above, we used learning rate scaling η ∝ √(B/B_ref). So it's a matter of ~10 lines of math to derive that to keep T_epoch constant, we need: +# λ = λ_ref · √(B/B_ref) · (D_ref/D) +# Note that these papers study AdamW, *not* Muon. We are blindly following AdamW theory for scaling hoping it ~works for Muon too. +weight_decay_scaled = args.weight_decay * math.sqrt(total_batch_size / B_REF) * (D_REF / target_tokens) +if weight_decay_scaled != args.weight_decay: + print0(f"Scaling weight decay from {args.weight_decay:.6f} to {weight_decay_scaled:.6f} for depth {args.depth}") + # ----------------------------------------------------------------------------- # Initialize the Optimizer (combined MuonAdamW: Muon for matrix params, AdamW for rest) -adam_betas = (args.adam_beta1, args.adam_beta2) optimizer = model.setup_optimizer( + # AdamW hyperparameters unembedding_lr=args.unembedding_lr * batch_lr_scale, embedding_lr=args.embedding_lr * batch_lr_scale, + scalar_lr=args.scalar_lr * batch_lr_scale, + adam_betas=(args.adam_beta1, args.adam_beta2), + # Muon hyperparameters matrix_lr=args.matrix_lr * batch_lr_scale, weight_decay=weight_decay_scaled, - adam_betas=adam_betas, - scalar_lr=args.scalar_lr * batch_lr_scale, ) if resuming: @@ -317,9 +323,30 @@ build_val_loader = lambda: tokenizing_distributed_data_loader_bos_bestfit(tokeni x, y, dataloader_state_dict = next(train_loader) # kick off load of the very first batch of data # ----------------------------------------------------------------------------- -# Set up hyperparameter schedulers +# Calculate the number of iterations we will train for and set up the various schedulers -# Learning rate scheduler +# num_iterations: either it is given, or from target flops, or from target data:param ratio (in that order) +assert args.num_iterations > 0 or args.target_param_data_ratio > 0 or args.target_flops > 0 +if args.num_iterations > 0: + # Override num_iterations to a specific value if given + num_iterations = args.num_iterations + print0(f"Using user-provided number of iterations: {num_iterations:,}") +elif args.target_flops > 0: + # Calculate the number of iterations from the target flops (used in scaling laws analysis, e.g. runs/scaling_laws.sh) + num_iterations = round(args.target_flops / (num_flops_per_token * total_batch_size)) + print0(f"Calculated number of iterations from target FLOPs: {num_iterations:,}") +elif args.target_param_data_ratio > 0: + # Calculate the number of iterations from the target param data ratio (the most common use case) + num_iterations = target_tokens // total_batch_size + print0(f"Calculated number of iterations from target data:param ratio: {num_iterations:,}") +else: + raise ValueError("No training horizon specified") +total_tokens = total_batch_size * num_iterations # the actual number of tokens we will train for +print0(f"Total number of training tokens: {total_tokens:,}") +print0(f"Tokens : Scaling params ratio: {total_batch_size * num_iterations / num_scaling_params:.2f}") # e.g. Chinchilla was ~20 +print0(f"Total training FLOPs estimate: {num_flops_per_token * total_tokens:e}") + +# Learning rate schedule (linear warmup, constant, linear warmdown) def get_lr_multiplier(it): warmup_iters = round(args.warmup_ratio * num_iterations) warmdown_iters = round(args.warmdown_ratio * num_iterations) @@ -331,19 +358,20 @@ def get_lr_multiplier(it): progress = (num_iterations - it) / warmdown_iters return progress * 1.0 + (1 - progress) * args.final_lr_frac -# Momentum scheduler for Muon optimizer +# Momentum scheduler for Muon optimizer (warms up to 0.95 over the first 300 steps) def get_muon_momentum(it): frac = min(it / 300, 1) momentum = (1 - frac) * 0.85 + frac * 0.95 return momentum -# Weight decay scheduler for Muon optimizer (linear to zero over the course of training) +# Weight decay scheduler for Muon optimizer (linearly decays to zero over the course of training) def get_weight_decay(it): return weight_decay_scaled * (1 - it / num_iterations) # ----------------------------------------------------------------------------- -# Loop state (variables updated by the training loop) +# Training loop +# Loop state (variables updated by the training loop) if not resuming: step = 0 val_bpb = None # will be set if eval_every > 0 @@ -358,11 +386,19 @@ else: smooth_train_loss = loop_state["smooth_train_loss"] total_training_time = loop_state["total_training_time"] -# ----------------------------------------------------------------------------- -# Training loop +# Figure out the needed gradient accumulation micro-steps to reach the desired total batch size per step +tokens_per_fwdbwd = args.device_batch_size * args.max_seq_len # tokens per iteration for a single rank +world_tokens_per_fwdbwd = tokens_per_fwdbwd * ddp_world_size # total tokens per iteration for all ranks +assert total_batch_size % world_tokens_per_fwdbwd == 0 +grad_accum_steps = total_batch_size // world_tokens_per_fwdbwd +print0(f"Tokens / micro-batch / rank: {args.device_batch_size} x {args.max_seq_len} = {tokens_per_fwdbwd:,}") +print0(f"Tokens / micro-batch: {world_tokens_per_fwdbwd:,}") +print0(f"Total batch size {total_batch_size:,} => gradient accumulation steps: {grad_accum_steps}") + +# Go! while True: last_step = step == num_iterations # loop runs num_iterations+1 times so that we can eval/save at the end - flops_so_far = num_flops_per_token * args.total_batch_size * step + flops_so_far = num_flops_per_token * total_batch_size * step # once in a while: evaluate the val bpb (all ranks participate) if args.eval_every > 0 and (last_step or step % args.eval_every == 0): @@ -434,6 +470,7 @@ while True: "user_config": user_config, # inputs to the training script "device_batch_size": args.device_batch_size, "max_seq_len": args.max_seq_len, + "total_batch_size": total_batch_size, "dataloader_state_dict": dataloader_state_dict, "loop_state": { # all loop state (other than step) so that we can resume training "min_val_bpb": min_val_bpb, @@ -482,8 +519,8 @@ while True: smooth_train_loss = ema_beta * smooth_train_loss + (1 - ema_beta) * train_loss_f # EMA the training loss debiased_smooth_loss = smooth_train_loss / (1 - ema_beta**(step + 1)) # debias the EMA pct_done = 100 * step / num_iterations - tok_per_sec = int(args.total_batch_size / dt) - flops_per_sec = num_flops_per_token * args.total_batch_size / dt + tok_per_sec = int(total_batch_size / dt) + flops_per_sec = num_flops_per_token * total_batch_size / dt mfu = 100 * flops_per_sec / (gpu_peak_flops * ddp_world_size) if step > 10: total_training_time += dt # only count the time after the first 10 steps @@ -497,7 +534,7 @@ while True: else: eta_str = "" epoch = dataloader_state_dict["epoch"] - print0(f"step {step:05d}/{num_iterations:05d} ({pct_done:.2f}%) | loss: {debiased_smooth_loss:.6f} | lrm: {lrm:.2f} | dt: {dt * 1000:.2f}ms | tok/sec: {tok_per_sec:,} | mfu: {mfu:.2f} | epoch: {epoch} | total time: {total_training_time/60:.2f}m{eta_str}") + print0(f"step {step:05d}/{num_iterations:05d} ({pct_done:.2f}%) | loss: {debiased_smooth_loss:.6f} | lrm: {lrm:.2f} | dt: {dt * 1000:.2f}ms | tok/sec: {tok_per_sec:,} | bf16_mfu: {mfu:.2f} | epoch: {epoch} | total time: {total_training_time/60:.2f}m{eta_str}") if step % 100 == 0: log_data = { "step": step, @@ -541,7 +578,7 @@ get_report().log(section="Base model training", data=[ "Number of FLOPs per token": f"{num_flops_per_token:e}", "Calculated number of iterations": num_iterations, "Number of training tokens": total_tokens, - "Tokens : Scaling params ratio": args.total_batch_size * num_iterations / num_scaling_params, + "Tokens : Scaling params ratio": total_batch_size * num_iterations / num_scaling_params, "DDP world size": ddp_world_size, "warmup_ratio": args.warmup_ratio, "warmdown_ratio": args.warmdown_ratio, diff --git a/scripts/chat_sft.py b/scripts/chat_sft.py index 4c81f06..a783ed2 100644 --- a/scripts/chat_sft.py +++ b/scripts/chat_sft.py @@ -9,6 +9,7 @@ Or torchrun for training: torchrun --standalone --nproc_per_node=8 -m scripts.chat_sft -- --device-batch-size=16 """ +import gc import argparse import os os.environ["PYTORCH_ALLOC_CONF"] = "expandable_segments:True" @@ -16,12 +17,14 @@ import time import wandb import torch from contextlib import nullcontext -from nanochat.common import compute_init, compute_cleanup, print0, DummyWandb, get_base_dir, autodetect_device_type +from nanochat.common import compute_init, compute_cleanup, print0, DummyWandb, get_base_dir, autodetect_device_type, get_peak_flops from nanochat.tokenizer import get_token_bytes -from nanochat.checkpoint_manager import save_checkpoint +from nanochat.checkpoint_manager import save_checkpoint, load_model, load_optimizer_state from nanochat.loss_eval import evaluate_bpb -from nanochat.checkpoint_manager import load_model import torch.distributed as dist +from nanochat.flash_attention import HAS_FA3 +from nanochat.engine import Engine +from scripts.chat_eval import run_chat_eval from tasks.common import TaskMixture from tasks.gsm8k import GSM8K @@ -37,27 +40,33 @@ parser = argparse.ArgumentParser(description="Supervised fine-tuning (SFT) the m parser.add_argument("--run", type=str, default="dummy", help="wandb run name ('dummy' disables wandb logging)") # Runtime parser.add_argument("--device-type", type=str, default="", help="cuda|cpu|mps (empty = autodetect)") -parser.add_argument("--dtype", type=str, default="bfloat16", help="float32|bfloat16") # Model loading parser.add_argument("--model-tag", type=str, default=None, help="model tag to load from") parser.add_argument("--model-step", type=int, default=None, help="model step to load from") +parser.add_argument("--load-optimizer", type=int, default=1, help="warm-start optimizer from pretrained checkpoint (0=no, 1=yes)") # Training horizon parser.add_argument("--num-iterations", type=int, default=-1, help="number of optimization steps (-1 = full epoch)") -# Batch sizes -parser.add_argument("--max-seq-len", type=int, default=2048, help="max context length") -parser.add_argument("--device-batch-size", type=int, default=32, help="per-device batch size") -parser.add_argument("--total-batch-size", type=int, default=524288, help="total batch size in tokens") -# Optimization -parser.add_argument("--embedding-lr", type=float, default=0.3, help="learning rate for embedding parameters (Adam)") -parser.add_argument("--unembedding-lr", type=float, default=0.004, help="learning rate for unembedding parameters (Adam)") -parser.add_argument("--matrix-lr", type=float, default=0.02, help="learning rate for matrix parameters (Muon)") -parser.add_argument("--weight-decay", type=float, default=0.0, help="weight decay for embedding/unembedding parameters (Adam)") -parser.add_argument("--init-lr-frac", type=float, default=1.0, help="initial LR as fraction of base LR") +# Batch sizes (default: inherit from pretrained checkpoint) +parser.add_argument("--max-seq-len", type=int, default=None, help="max context length (default: inherit from pretrain)") +parser.add_argument("--device-batch-size", type=int, default=None, help="per-device batch size (default: inherit from pretrain)") +parser.add_argument("--total-batch-size", type=int, default=None, help="total batch size in tokens (default: inherit from pretrain)") +# Optimization (default: inherit from pretrained checkpoint) +parser.add_argument("--embedding-lr", type=float, default=None, help="learning rate for embedding parameters (Adam) (default: inherit from pretrain)") +parser.add_argument("--unembedding-lr", type=float, default=None, help="learning rate for unembedding parameters (Adam) (default: inherit from pretrain)") +parser.add_argument("--matrix-lr", type=float, default=None, help="learning rate for matrix parameters (Muon) (default: inherit from pretrain)") +parser.add_argument("--init-lr-frac", type=float, default=0.8, help="initial LR as fraction of base LR") +parser.add_argument("--warmup-ratio", type=float, default=0.0, help="ratio of iterations for LR warmup") +parser.add_argument("--warmdown-ratio", type=float, default=0.5, help="ratio of iterations for LR warmdown") +parser.add_argument("--final-lr-frac", type=float, default=0.0, help="final LR as fraction of initial LR") # Evaluation -parser.add_argument("--eval-every", type=int, default=150, help="evaluate val bpb every N steps (-1 = disable)") -parser.add_argument("--eval-tokens", type=int, default=20*524288, help="number of tokens to evaluate val loss on") -# Output -parser.add_argument("--dry-run", action="store_true", help="log to wandb but skip checkpoints/report") +parser.add_argument("--eval-every", type=int, default=200, help="evaluate val bpb every N steps (-1 = disable)") +parser.add_argument("--eval-tokens", type=int, default=40*524288, help="number of tokens to evaluate val loss on") +parser.add_argument("--chatcore-every", type=int, default=200, help="evaluate ChatCORE metric every N steps (-1 = disable)") +parser.add_argument("--chatcore-max-cat", type=int, default=-1, help="max problems per categorical task for ChatCORE") +parser.add_argument("--chatcore-max-sample", type=int, default=24, help="max problems per generative task for ChatCORE") +# Data mixture +parser.add_argument("--mmlu-epochs", type=int, default=3, help="number of epochs of MMLU in training mixture (teaches Multiple Choice)") +parser.add_argument("--gsm8k-epochs", type=int, default=4, help="number of epochs of GSM8K in training mixture (teaches Math and Tool Use)") args = parser.parse_args() user_config = vars(args).copy() # ----------------------------------------------------------------------------- @@ -66,20 +75,48 @@ user_config = vars(args).copy() device_type = autodetect_device_type() if args.device_type == "" else args.device_type ddp, ddp_rank, ddp_local_rank, ddp_world_size, device = compute_init(device_type) master_process = ddp_rank == 0 -ptdtype = torch.float32 if args.dtype == 'float32' else torch.bfloat16 -autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype) if device_type == "cuda" else nullcontext() +autocast_ctx = torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16) if device_type == "cuda" else nullcontext() synchronize = torch.cuda.synchronize if device_type == "cuda" else lambda: None get_max_memory = torch.cuda.max_memory_allocated if device_type == "cuda" else lambda: 0 +if device_type == "cuda": + gpu_device_name = torch.cuda.get_device_name(0) + gpu_peak_flops = get_peak_flops(gpu_device_name) + print0(f"GPU: {gpu_device_name} | Peak FLOPS (BF16): {gpu_peak_flops:.2e}") +else: + gpu_peak_flops = float('inf') # MFU not meaningful for CPU/MPS # wandb logging init use_dummy_wandb = args.run == "dummy" or not master_process wandb_run = DummyWandb() if use_dummy_wandb else wandb.init(project="nanochat-sft", name=args.run, config=user_config) +# Flash Attention status +if not HAS_FA3: + print0("WARNING: Flash Attention 3 not available, using PyTorch SDPA fallback. Training will be less efficient.") + # Load the model and tokenizer model, tokenizer, meta = load_model("base", device, phase="train", model_tag=args.model_tag, step=args.model_step) -pretrain_batch_size = meta.get("device_batch_size", None) -if pretrain_batch_size is not None and args.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?") + +# Inherit training hyperparameters from pretrained checkpoint (None = inherit, explicit value = override) +pretrain_user_config = meta.get("user_config", {}) +for name, fallback, source in [ + ("max_seq_len", 2048, meta), + ("device_batch_size", 32, meta), + ("total_batch_size", 524288, meta), + ("embedding_lr", 0.3, pretrain_user_config), + ("unembedding_lr", 0.004, pretrain_user_config), + ("matrix_lr", 0.02, pretrain_user_config), +]: + arg_val = getattr(args, name) + pretrain_val = source.get(name) + if arg_val is None: + resolved = pretrain_val if pretrain_val is not None else fallback + setattr(args, name, resolved) + print0(f"Inherited {name}={resolved} from pretrained checkpoint") + elif pretrain_val is not None and arg_val != pretrain_val: + print0(f"NOTE: --{name.replace('_', '-')}={arg_val} overrides pretrained value of {pretrain_val}") + else: + print0(f"Using {name}={arg_val}") + orig_model = model model = torch.compile(model, dynamic=False) depth = model.config.n_layer @@ -94,25 +131,44 @@ print0(f"Total batch size {args.total_batch_size:,} => gradient accumulation ste token_bytes = get_token_bytes(device=device) # Initialize the Optimizer (combined MuonAdamW: Muon for matrix params, AdamW for rest) -optimizer = model.setup_optimizer(unembedding_lr=args.unembedding_lr, embedding_lr=args.embedding_lr, matrix_lr=args.matrix_lr, weight_decay=args.weight_decay) +# Note that pretraining ramps weight_decay to zero by end of pretraining, so SFT continues with zero +optimizer = model.setup_optimizer(unembedding_lr=args.unembedding_lr, embedding_lr=args.embedding_lr, matrix_lr=args.matrix_lr, weight_decay=0.0) + +# Optionally warm-start optimizer from pretrained checkpoint (momentum buffers etc.) +# Note: load_state_dict overwrites param_group metadata (LRs, betas, etc.) with the +# pretrained values. Since pretraining warmdown brings LRs to ~0, we must save and +# restore our fresh SFT LRs after loading. +base_dir = get_base_dir() +if args.load_optimizer: + optimizer_data = load_optimizer_state("base", device, rank=ddp_rank, model_tag=args.model_tag, step=args.model_step) + if optimizer_data is not None: + base_lrs = [group["lr"] for group in optimizer.param_groups] + optimizer.load_state_dict(optimizer_data) + del optimizer_data + for group, base_lr in zip(optimizer.param_groups, base_lrs): + group["lr"] = base_lr + print0("Loaded optimizer state from pretrained checkpoint (momentum buffers only, LRs reset)") + else: + print0("WARNING: optimizer checkpoint not found, starting with fresh optimizer (slightly worse)") + # Override the initial learning rate as a fraction of the base learning rate for group in optimizer.param_groups: group["lr"] = group["lr"] * args.init_lr_frac group["initial_lr"] = group["lr"] # SFT data mixture and DataLoader -base_dir = get_base_dir() identity_conversations_filepath = os.path.join(base_dir, "identity_conversations.jsonl") -train_dataset = TaskMixture([ +train_tasks = [ SmolTalk(split="train"), # 460K rows of general conversations - MMLU(subset="auxiliary_train", split="train"), # 100K rows of multiple choice problems drawn from ARC, MC_TEST, OBQA, RACE - GSM8K(subset="main", split="train"), # 8K rows teaching simple math and (calculator) tool use - GSM8K(subset="main", split="train"), # 2 epochs of GSM8K CustomJSON(filepath=identity_conversations_filepath), # 1000 rows of synthetic identity conversations - CustomJSON(filepath=identity_conversations_filepath), # let's do 2 epochs of these + CustomJSON(filepath=identity_conversations_filepath), # 2 epochs of these + *[MMLU(subset="auxiliary_train", split="train") for _ in range(args.mmlu_epochs)], # 100K rows per epoch + *[GSM8K(subset="main", split="train") for _ in range(args.gsm8k_epochs)], # 8K rows per epoch SimpleSpelling(size=200000, split="train"), # 200K rows of Simple Spelling (e.g. spell the word 'apple') SpellingBee(size=80000, split="train"), # 80K rows of Spelling Bee (e.g. how many 'r' are in 'strawberry'?) -]) # total: 460K + 100K + 16K + 200K + 80K = 856K rows +] +train_dataset = TaskMixture(train_tasks) +print0(f"Training mixture: {len(train_dataset):,} rows (MMLU x{args.mmlu_epochs}, GSM8K x{args.gsm8k_epochs})") val_dataset = TaskMixture([ SmolTalk(split="test"), # 24K rows in test set MMLU(subset="all", split="test", stop=5200), # 14K rows in test set, use only 5.2K to match the train ratios @@ -236,10 +292,17 @@ train_loader = sft_data_generator_bos_bestfit("train") build_val_loader = lambda: sft_data_generator_bos_bestfit("val") progress = 0 # will go from 0 to 1 over the course of the epoch -# Learning rate scheduler +# Learning rate schedule (linear warmup, constant, linear warmdown) +# Same shape as base_train but uses progress (0→1) instead of absolute step counts, +# because SFT doesn't always know num_iterations in advance (dataset-driven stopping). def get_lr_multiplier(progress): - # first 80% of training: no decay, then linearly ramp down to 0. - return 1 if progress < 0.8 else 1 - (progress - 0.8) / 0.2 + if progress < args.warmup_ratio: + return (progress + 1e-8) / args.warmup_ratio + elif progress <= 1.0 - args.warmdown_ratio: + return 1.0 + else: + decay = (progress - (1.0 - args.warmdown_ratio)) / args.warmdown_ratio + return (1 - decay) * 1.0 + decay * args.final_lr_frac # Momentum scheduler for Muon optimizer def get_muon_momentum(it): @@ -282,8 +345,44 @@ while True: }) model.train() - # save checkpoint at the end of the run (only on master process) - if master_process and last_step and not args.dry_run: + # once in a while: estimate the ChatCORE metric (all ranks participate) + # use the original uncompiled model because the inputs keep changing shape + chatcore_results = {} + if args.chatcore_every > 0 and (last_step or (step > 0 and step % args.chatcore_every == 0)): + model.eval() + engine = Engine(orig_model, tokenizer) + all_tasks = ['ARC-Easy', 'ARC-Challenge', 'MMLU', 'GSM8K', 'HumanEval', 'SpellingBee'] + categorical_tasks = {'ARC-Easy', 'ARC-Challenge', 'MMLU'} + baseline_accuracies = { + 'ARC-Easy': 0.25, 'ARC-Challenge': 0.25, 'MMLU': 0.25, + 'GSM8K': 0.0, 'HumanEval': 0.0, 'SpellingBee': 0.0, + } + task_results = {} + for task_name in all_tasks: + limit = args.chatcore_max_cat if task_name in categorical_tasks else args.chatcore_max_sample + max_problems = None if limit < 0 else limit # -1 means no limit + with autocast_ctx: + acc = run_chat_eval(task_name, orig_model, tokenizer, engine, + batch_size=args.device_batch_size, max_problems=max_problems) + task_results[task_name] = acc + print0(f" {task_name}: {100*acc:.2f}%") + # Compute ChatCORE metrics (mean centered accuracy, ranges from 0=random to 1=perfect) + def centered_mean(tasks): + return sum((task_results[t] - baseline_accuracies[t]) / (1.0 - baseline_accuracies[t]) for t in tasks) / len(tasks) + chatcore = centered_mean(all_tasks) + chatcore_cat = centered_mean(categorical_tasks) + print0(f"Step {step:05d} | ChatCORE: {chatcore:.4f} | ChatCORE_cat: {chatcore_cat:.4f}") + wandb_run.log({ + "step": step, + "total_training_flops": flops_so_far, + "chatcore_metric": chatcore, + "chatcore_cat": chatcore_cat, + **{f"chatcore/{task_name}": acc for task_name, acc in task_results.items()}, + }) + model.train() + + # save checkpoint at the end of the run (all ranks participate so each saves its optimizer shard) + if last_step: output_dirname = args.model_tag if args.model_tag else f"d{depth}" # e.g. d12 checkpoint_dir = os.path.join(base_dir, "chatsft_checkpoints", output_dirname) save_checkpoint( @@ -304,7 +403,8 @@ while True: "window_pattern": model.config.window_pattern, }, "user_config": user_config, # inputs to the training script - } + }, + rank=ddp_rank, ) if last_step: @@ -346,8 +446,7 @@ while True: pct_done = 100 * progress tok_per_sec = int(args.total_batch_size / dt) flops_per_sec = num_flops_per_token * args.total_batch_size / dt - promised_flops_per_sec_h100 = 989e12 * ddp_world_size # bfloat16 H100 SXM and without 2:4 sparsity - mfu = 100 * flops_per_sec / promised_flops_per_sec_h100 # in % + mfu = 100 * flops_per_sec / (gpu_peak_flops * ddp_world_size) if step > 10: total_training_time += dt # only count the time after the first 10 steps print0(f"step {step:05d} ({pct_done:.2f}%) | loss: {debiased_smooth_loss:.6f} | lrm: {lrm:.2f} | dt: {dt * 1000:.2f}ms | tok/sec: {tok_per_sec:,} | mfu: {mfu:.2f} | epoch: {current_epoch} | total time: {total_training_time/60:.2f}m") @@ -364,24 +463,32 @@ while True: "train/epoch": current_epoch, }) + # The garbage collector spends ~500ms scanning for cycles quite frequently. + # We manually manage it to avoid these pauses during training. + if step == 1: + gc.collect() # manually collect a lot of garbage from setup + gc.freeze() # freeze all currently surviving objects and exclude them from GC + gc.disable() # disable GC entirely except: + elif step % 5000 == 0: # every 5000 steps... + gc.collect() # manually collect, just to be safe for very long runs + # print a few more stats print0(f"Peak memory usage: {get_max_memory() / 1024 / 1024:.2f}MiB") print0(f"Total training time: {total_training_time/60:.2f}m") print0(f"Minimum validation bpb: {min_val_bpb:.4f}") # Log to report -if not args.dry_run: - from nanochat.report import get_report - get_report().log(section="SFT", data=[ - user_config, # CLI args - { # stats about the training setup - "Number of iterations": step, - "DDP world size": ddp_world_size, - }, - { # stats about training outcomes - "Minimum validation bpb": min_val_bpb, - } - ]) +from nanochat.report import get_report +get_report().log(section="SFT", data=[ + user_config, # CLI args + { # stats about the training setup + "Number of iterations": step, + "DDP world size": ddp_world_size, + }, + { # stats about training outcomes + "Minimum validation bpb": min_val_bpb, + } +]) # cleanup wandb_run.finish() # wandb run finish diff --git a/tests/test_engine.py b/tests/test_engine.py index 0159111..784ffcb 100644 --- a/tests/test_engine.py +++ b/tests/test_engine.py @@ -31,7 +31,7 @@ class MockModel: def __init__(self, vocab_size=262): # 256 bytes + 6 special tokens self.vocab_size = vocab_size self.config = MockConfig() - self._device = "cpu" + self._device = torch.device("cpu") def get_device(self): return self._device diff --git a/uv.lock b/uv.lock index e5fc97f..bbc9519 100644 --- a/uv.lock +++ b/uv.lock @@ -1509,7 +1509,6 @@ dependencies = [ { name = "torch", version = "2.9.1", source = { registry = "https://pypi.org/simple" }, marker = "(extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu') or (extra != 'extra-8-nanochat-cpu' and extra != 'extra-8-nanochat-gpu')" }, { name = "torch", version = "2.9.1+cpu", source = { registry = "https://download.pytorch.org/whl/cpu" }, marker = "(sys_platform != 'darwin' and extra == 'extra-8-nanochat-cpu') or (extra == 'extra-8-nanochat-cpu' and extra == 'extra-8-nanochat-gpu')" }, { name = "torch", version = "2.9.1+cu128", source = { registry = "https://download.pytorch.org/whl/cu128" }, marker = "extra == 'extra-8-nanochat-gpu'" }, - { name = "torchao" }, { name = "transformers" }, { name = "uvicorn" }, { name = "wandb" }, @@ -1549,7 +1548,6 @@ requires-dist = [ { name = "torch", specifier = "==2.9.1" }, { name = "torch", marker = "extra == 'cpu'", specifier = "==2.9.1", index = "https://download.pytorch.org/whl/cpu", conflict = { package = "nanochat", extra = "cpu" } }, { name = "torch", marker = "extra == 'gpu'", specifier = "==2.9.1", index = "https://download.pytorch.org/whl/cu128", conflict = { package = "nanochat", extra = "gpu" } }, - { name = "torchao", specifier = "==0.15.0" }, { name = "transformers", specifier = ">=4.57.3" }, { name = "uvicorn", specifier = ">=0.36.0" }, { name = "wandb", specifier = ">=0.21.3" }, @@ -3184,15 +3182,6 @@ wheels = [ { url = "https://download.pytorch.org/whl/cu128/torch-2.9.1%2Bcu128-cp314-cp314t-win_amd64.whl", hash = "sha256:0c784b600959ec70ee01cb23e8bc870a0e0475af30378ff5e39f4abed8b7c1cc" }, ] -[[package]] -name = "torchao" -version = "0.15.0" -source = { registry = "https://pypi.org/simple" } -wheels = [ - { url = "https://files.pythonhosted.org/packages/57/2d/472b9362dceae05a4599e2b94f86e69a29c0e20964a6af84f34f6ead5938/torchao-0.15.0-cp310-abi3-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:1cbe813201314ba6329a650a76944502f3e8ec4b1b44523f3f48676810d8d1f6", size = 7163930, upload-time = "2025-12-18T23:14:41.876Z" }, - { url = "https://files.pythonhosted.org/packages/f6/3b/6b9d5618720f63dbc2e2509cd6b57aae9c0d61b738d1d2172f4d5d9efaab/torchao-0.15.0-py3-none-any.whl", hash = "sha256:3f3812676048ef8a2a0e9d492d12d8971ba7a7ebb16f54aa56f690414e130d2c", size = 1080679, upload-time = "2025-12-18T23:14:43.807Z" }, -] - [[package]] name = "tornado" version = "6.5.4"