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>
## Problem
When running SFT with small device-batch-size (≤8), fully-masked micro-batches
cause NaN loss from step 1, corrupting gradients permanently. This happens when
a micro-batch contains only 'User' tokens (all targets=-1), especially common
with small batch sizes on consumer GPUs.
Root cause: torch.nn.functional.cross_entropy with reduction='mean' returns NaN
when all labels are -1 (division by zero in mean computation).
## Solution
Added validation in the training loop to detect and skip fully-masked batches:
- Check (y != -1).any() before computing loss
- Skip backward() for batches with no valid targets (zero gradient contribution)
- Track skipped batches and warn user if >5% in first 100 steps
- Log skipped batches as loss=0 for transparency
## Testing
- Added comprehensive test suite (test_sft_masked_batches.py)
- Tests cover: fully masked, partially masked, and unmasked batches
- Documents cross_entropy behavior with ignore_index=-1
- Validates the fix logic
## Impact
- Fixes#590: NaN loss with small batch sizes
- No performance impact for normal batches
- Helps users on consumer GPUs (RTX 3060, etc.)
- Prevents silent gradient corruption
Resolves#590
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
