W1 needs the GPT to consume Whisper-projector outputs as a prefix of "soft
tokens" sitting in front of the text token embeddings (LLaVA-style). The
change is intentionally minimal:
- forward() takes a new keyword-only arg `audio_features` of shape
(B, T_a, n_embd). They must already be projected to n_embd by the caller
(Projector lives in nanochat.audio, kept out of GPT itself).
- The audio rows are normed (matches the post-wte norm convention) and
concatenated *after* smear so smear stays a strictly text-side op (its
prev-token semantics aren't defined for soft tokens, and revisiting that
belongs to a later phase).
- Rotary embeddings are re-sliced for T_a + T_text. Audio gets positions
0..T_a-1, text 0-shifts to T_a..T_full-1. The 10× over-allocated rotary
cache in __init__ already covers this.
- value_embeds lookup uses an idx padded with 0 for audio positions. They
feed the v residual but the gate (`ve_gate`) is input-dependent and will
learn to suppress the dummy rows; for W1 smoke this is fine.
- targets are auto-padded with -1 (ignore_index) over audio positions so
the LM is only graded on text predictions.
Not yet supported: audio_features with kv_cache. The KV-cache path is a
prefill+decode protocol that would need its own audio-aware semantics; W1
runs train-style forwards only, so we just assert.
- pyproject.toml + uv.lock: pytorch-cu128/cpu indexes → mirror.sjtu.edu.cn
(aliyun lacks 2.9.1, sjtu has it)
- nanochat/dataset.py: climbmix BASE_URL → hf-mirror.com
For ailab (CN, RTX 5090) where direct pytorch.org and huggingface.co
are unreachable. Override at uv-sync time with UV_DEFAULT_INDEX env.
When swapping Float8Linear to Linear in disable_fp8 context manager,
using device=fp8_module.weight.device directly allocates new tensors
on GPU, causing unnecessary VRAM spike (~1GB for large models).
This fix uses device='meta' to avoid physical memory allocation,
then swaps in the weight tensor reference. This eliminates the
unnecessary VRAM spike during evaluation phase.
Fixes issue #592
Co-authored-by: RoomWithOutRoof <roomwithoutroof@sparklab.ai>
The bf16 cast is intentional for speed on Hopper+ GPUs, but should be
skipped on other platforms rather than blindly applied. fp16 is unstable
here due to its limited exponent range, and fp32 platforms don't benefit
from the cast. Now: bf16 when COMPUTE_DTYPE is bf16, no cast otherwise.
Inspired by PR #667.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New architectural features:
- Smear: mix previous token embedding into current position via learned
gate, providing cheap bigram-like info (works in training + KV cache)
- Backout: subtract learned fraction of mid-layer residual before logit
projection to remove low-level features
Hyperparameter tuning:
- Muon momentum warmdown 0.97→0.90 during LR warmdown phase
- Non-uniform per-layer init: resid_lambdas 1.15→1.05, x0_lambdas 0.20→0.05
- c_fc init scale 0.4x, QK norm scale 1.2, sliding window seq_len/4
- Speedrun data:params ratio reduced to 8
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* printing steps count
* adding reply only loss for chat
* using the mask by render_conversation function of tokeniser
* undoing some changes
* putting back the comment which got removed accidently, no functionality change
Fam Zheng
fam