Skip to content

Support DSV4 IndexCache training, distillation, and recompute #1429

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
chenbong wants to merge 10 commits into
base: release/0.3
Choose a base branch
from
Draft

Support DSV4 IndexCache training, distillation, and recompute #1429

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
83c55bf
Fix IndexCache reuse producer debug
Jul 6, 2026
750a7f2
Support IndexCache FSF PP state handoff
Jul 6, 2026
e9add54
Support IndexCache training safeguards
Jul 7, 2026
15dbda3
Support IndexCache reuse recompute state guards
Jul 8, 2026
f228558
Support IndexCache reuse recompute PP flow
Jul 8, 2026
aa8ffff
Support IndexCache distill recompute training
Jul 8, 2026
4a57387
Match hash router table checkpoint dtype
Jul 11, 2026
e3d26fc
Fix IndexCache PP missing gradients
Jul 12, 2026
7ad0c70
Fix IndexCache outer pipeline gradients
Jul 12, 2026
10fd153
Preserve released IndexCache PP metadata
Jul 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions src/paddlefleet/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,10 @@

# Re-export from Paddle for backward compatibility (PaddleFormers imports these from paddlefleet)

from .indexcache_pp_runtime_patch import apply_indexcache_pp_runtime_patch

apply_indexcache_pp_runtime_patch()

from . import (
parallel_state as parallel_state,
training as training,
Expand Down
2 changes: 2 additions & 0 deletions src/paddlefleet/fusions/csa_sparse_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -107,6 +107,7 @@ def forward(
ctx.query_shape = (b, sq, np_heads, hn)
ctx.softmax_scale = float(softmax_scale)
ctx.attn_sink_dtype = attn_sink.dtype
ctx.attn_sink_stop_gradient = bool(attn_sink.stop_gradient)
ctx.backend = backend

query, kv_full, attn_sink, topk_idxs = prepare_inputs(
Expand Down Expand Up @@ -195,6 +196,7 @@ def backward(ctx, grad_output):
ctx.attn_sink_dtype
)

d_attn_sink = None if ctx.attn_sink_stop_gradient else d_attn_sink
return (dq, dkv, d_attn_sink, None)


Expand Down
Loading