Found while gating #1458 (PR #1461). Filed in flow; it is not that change's
defect and it is deliberately not repaired there, because it is the same class of
bound #1461 exists to remove and it needs its own derivation.
What it is
tests/vllm/models/test_muse_glimmer_text.cpp:532
diff is the whole text forward against the independently transcribed fp32
reference. The constant is the W1 measurement rounded up: 3a54c4b7d's own body
quotes "max|diff| 1.21e-4 on logits of max 4.88e-2", and 5e-4 is that with 4.1x
of headroom. There is no derivation beside it, and it is the last remaining
member of that class in this file — #1461 replaces the other one
(bdiff <= 1e-5) in the same case.
Why it matters now
4712dac40 narrowed act(gate) to the input dtype, which is upstream's
polarity, and grew this envelope 2.8x:
| kernel |
diff |
ratio to the 5e-4 bound |
before 4712dac40 |
1.21073e-04 |
0.242 |
at aeba0de6f and after |
3.43494e-04 |
0.687 |
One correct kernel change consumed 45 points of a 76-point margin. The next
legitimate change to any op on this path — an accumulation order, a store width,
a fused chain that rounds where the unfused one did not — reds a green suite and
presents as a defect in whatever is in flight. That is exactly how #1458 was
found.
What a repair owes
Not a bigger constant. Either a measured floor of the kind
test_ltx2_text_encoder propagates from the oracle's own f32-vs-bf16 spread, or
a bound derived from bf16's representable precision at this model's depth and
shapes, written down beside the number. If neither can be produced, the finding
is that this comparison cannot carry a tight band and the case's discriminating
power has to come from somewhere else — the mechanism property tests below it
already move the output by O(1), so the band is not what makes them work.
Measured at aeba0de6f, CPU-only Release (so NDEBUG), x86_64.
Owner: row MODEL-MUSE-GLIMMER, spec .agents/specs/muse-glimmer.md, under ## Owed.
Found while gating #1458 (PR #1461). Filed in flow; it is not that change's
defect and it is deliberately not repaired there, because it is the same class of
bound #1461 exists to remove and it needs its own derivation.
What it is
tests/vllm/models/test_muse_glimmer_text.cpp:532diffis the whole text forward against the independently transcribed fp32reference. The constant is the W1 measurement rounded up:
3a54c4b7d's own bodyquotes "max|diff| 1.21e-4 on logits of max 4.88e-2", and
5e-4is that with 4.1xof headroom. There is no derivation beside it, and it is the last remaining
member of that class in this file — #1461 replaces the other one
(
bdiff <= 1e-5) in the same case.Why it matters now
4712dac40narrowedact(gate)to the input dtype, which is upstream'spolarity, and grew this envelope 2.8x:
diff5e-4bound4712dac40aeba0de6fand afterOne correct kernel change consumed 45 points of a 76-point margin. The next
legitimate change to any op on this path — an accumulation order, a store width,
a fused chain that rounds where the unfused one did not — reds a green suite and
presents as a defect in whatever is in flight. That is exactly how #1458 was
found.
What a repair owes
Not a bigger constant. Either a measured floor of the kind
test_ltx2_text_encoderpropagates from the oracle's own f32-vs-bf16 spread, ora bound derived from bf16's representable precision at this model's depth and
shapes, written down beside the number. If neither can be produced, the finding
is that this comparison cannot carry a tight band and the case's discriminating
power has to come from somewhere else — the mechanism property tests below it
already move the output by O(1), so the band is not what makes them work.
Measured at
aeba0de6f, CPU-only Release (so NDEBUG), x86_64.Owner: row
MODEL-MUSE-GLIMMER, spec.agents/specs/muse-glimmer.md, under## Owed.