FA-98846 / Gradient accumulation weighting / Open access
Unequal valid-token batches receive equal token-mean weight · case 01
Unequal valid-token batches receive equal token-mean weight.
ROOT CAUSE
The broken path averages segment means even when segments contain different numbers of valid tokens.
THE FAILURE
The broken path averages segment means even when segments contain different numbers of valid tokens.
Unsuccessful approach: A microbatch mean still gives short and long token segments equal influence.
Case contract
Each sum contains only valid-token contributions. Divide the combined sum by all valid tokens; no valid tokens yields zero.
Why this case matters
A deterministic toy model isolates one local accumulation contract.
1 / The failure
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(valid_sums, valid_counts):
return sum(s / max(1,c) for s,c in zip(valid_sums,valid_counts)) / max(1,len(valid_sums))
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('unequal token lengths',solve([2*N,12*N],[N,3*N]),3.5)
check('equal lengths',solve([2*N,4*N],[N,N]),3)
check('single segment',solve([5*N],[N]),5)
check('all empty',solve([0],[0]),0)
check('empty list',solve([],[]),0)
check('zero segment',solve([0,9*N],[N,3*N]),2.25)
check('three segments',solve([N,6*N,8*N],[1,2,4]),15*N/7)
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)
| Boundary fixture | Actual | Expected | Outcome |
|---|---|---|---|
| unequal token lengths | 3.0 | 3.5 | Failed |
| equal lengths | 3.0 | 3 | Passed |
| single segment | 5.0 | 5 | Passed |
| all empty | 0.0 | 0 | Passed |
| empty list | 0.0 | 0 | Passed |
| zero segment | 1.5 | 2.25 | Failed |
| three segments | 2.0 | 2.142857142857143 | Failed |
SHA-256 / 67ed16afade97665ecf88f14b5b66b1c67023daad54284f97399d7ec129b05fd
2 / The unsuccessful fix
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(valid_sums, valid_counts):
return sum(valid_sums) / max(1,len(valid_sums))
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('unequal token lengths',solve([2*N,12*N],[N,3*N]),3.5)
check('equal lengths',solve([2*N,4*N],[N,N]),3)
check('single segment',solve([5*N],[N]),5)
check('all empty',solve([0],[0]),0)
check('empty list',solve([],[]),0)
check('zero segment',solve([0,9*N],[N,3*N]),2.25)
check('three segments',solve([N,6*N,8*N],[1,2,4]),15*N/7)
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)
| Boundary fixture | Actual | Expected | Outcome |
|---|---|---|---|
| unequal token lengths | 7.0 | 3.5 | Failed |
| equal lengths | 3.0 | 3 | Passed |
| single segment | 5.0 | 5 | Passed |
| all empty | 0.0 | 0 | Passed |
| empty list | 0.0 | 0 | Passed |
| zero segment | 4.5 | 2.25 | Failed |
| three segments | 5.0 | 2.142857142857143 | Failed |
SHA-256 / 8058abeb2959de7058ad9746655f6e729f762877a521a4c56bf479a8bf47af87
HELD IN THE MEMBER ARCHIVE
The verified repair and its recorded checks are member-only.
This mechanism has 7 recorded checks per implementation. The open-access tier publishes the failure and the unsuccessful fix; the repaired source that passes every check, and the observations that prove it, are available to members.
Every case sharing this mechanism uses the same contract and the same repair, so this one record is held back for all of them.
Member access is invitation-based. Sign in with your invited account to inspect the repair.
Sign in to the archive ↗Verification & scope
Inputs are scalar or short-list summaries and do not model autodiff, devices, or an optimizer. This reproducer isolates one failure mechanism. Results cover the supplied fixtures. Variants within a family share a test contract and should remain grouped when constructing evaluation splits. Related mechanisms with a shared evaluation_group must also remain together; these controlled models are not independent production incidents.
Observations recorded using Python 3.12.14 at 2026-09-30T16:48:17.606517+00:00.
Case digest / 4ff21167630a140229abc357ae31c96a91e5c5437ccf2b28f4f0b28fde241baf