[CI] Ignore ruff D209, rejoin the docstrings it split, and silence incompatible-rule warnings (#57212)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-16 10:57:24 -07:00
committed by GitHub
co-authored by Claude Opus 5
parent 1085b64425
commit 651a88c09a
693 changed files with 2088 additions and 4171 deletions
+2 -4
View File
@@ -60,8 +60,7 @@ def dmel_embed_sum_norm(
eps: float,
) -> torch.Tensor:
"""``rmsnorm(sum_b weight[b * VOCAB + idx[:, b]])`` in one launch (no
[T, NB, D] intermediate).
"""
[T, NB, D] intermediate)."""
T, nb = dmel_idx.shape
D = weight.shape[1]
assert weight.shape[0] % nb == 0
@@ -152,8 +151,7 @@ def rmsnorm_gelu(
) -> torch.Tensor:
"""Fused ``gelu(rmsnorm(x))`` (or plain rmsnorm); multiple rows per block
when D is small. With ``fold``, x must be [N, T, H, W, D] and the output
comes back as ``fold_timespace_to_depth(result, *fold)``.
"""
comes back as ``fold_timespace_to_depth(result, *fold)``."""
D = x.shape[-1]
flat = x.reshape(-1, D)
assert flat.stride(1) == 1 and flat.stride(0) == D
+2 -4
View File
@@ -238,8 +238,7 @@ def embed_rmsnorm(
embedding-table gather.
``chain_weight`` additionally emits ``rmsnorm(out, chain_weight)`` (the
first decoder layer's pre-attention norm) as a second output, still one
launch. Bit-exact vs the unfused module sequence.
"""
launch. Bit-exact vs the unfused module sequence."""
ids = input_ids.view(-1)
(T,) = ids.shape
n = embed_table.shape[1]
@@ -333,8 +332,7 @@ def embed_dual_rmsnorm_cat(
(draft decode steps) or precomputed ``embeds`` ([T, N], the target-merged
multimodal embeddings at draft prefill); ``pre_norm_weight`` chains the
backbone embed_norm in front of the depth embed_norm (bit-exact vs the
unfused sequence). The concat copies collapse into direct writes.
"""
unfused sequence). The concat copies collapse into direct writes."""
T, n = hidden.shape
if embeds is not None:
assert embeds.shape == hidden.shape