Skip to content

th#1043: forced group-fit fp8 survives the bf16-resident upcast (0.48.1)#375

Merged
PaulFidika merged 1 commit into
masterfrom
release/th1043-resident-upgrade-guard
Jul 23, 2026
Merged

th#1043: forced group-fit fp8 survives the bf16-resident upcast (0.48.1)#375
PaulFidika merged 1 commit into
masterfrom
release/th1043-resident-upgrade-guard

Conversation

@PaulFidika

Copy link
Copy Markdown
Contributor

Summary

  • Live 0.48.0 verification (pod 4xh4m999n26u5f, A100-80GB, bare-bf16 qwen-image bindings) showed the th#1043 joint fp8 forcing being silently defeated one layer deeper: load_from_pretrained's gw#534 bf16-resident upcast check evaluates a SINGLE lane against current free VRAM, saw the first lane fit alone, and upgraded it back to full bf16 residency (in_vram 43.2GB) — re-starving the sibling lane into the offload placement the shared-component invariant refuses.
  • Fix: force_storage_dtype (the joint group decision) now passes allow_bf16_resident_upgrade=False — a forced fp8 lane never locally upgrades; the headroom belongs to the group.
  • Version 0.48.1 + changelog in the same train.

Test plan

  • New regression test: with bf16_resident_fits returning True, an unforced fp8 load still upgrades (gw#534 preserved) while a forced one keeps fp8 hooks armed (real tiny diffusers pipeline, CPU)
  • mypy + ruff clean; quantization-lane suite green
  • Live qwen-image bf16 generate on the rebuilt image (in progress, the actual th#1043 acceptance)

Live 0.48.0 pod showed the joint fp8 forcing being silently un-done by
load_from_pretrained's single-lane resident-upcast check (gw#534 rung 2):
the first lane fit current free VRAM alone, upgraded back to bf16, and
re-starved its sibling into the refused offload placement. A forced
storage dtype now disables the local upgrade — the headroom belongs to
the group, not the first lane to load.
@PaulFidika
PaulFidika merged commit ef024a0 into master Jul 23, 2026
1 check passed
@PaulFidika
PaulFidika deleted the release/th1043-resident-upgrade-guard branch July 23, 2026 18:51
PaulFidika added a commit that referenced this pull request Jul 23, 2026
….1) (#375)

Live 0.48.0 pod showed the joint fp8 forcing being silently un-done by
load_from_pretrained's single-lane resident-upcast check (gw#534 rung 2):
the first lane fit current free VRAM alone, upgraded back to bf16, and
re-starved its sibling into the refused offload placement. A forced
storage dtype now disables the local upgrade — the headroom belongs to
the group, not the first lane to load.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant