JAXBench: Fix no-op exploit in 19 KernelBench baselines: zero-init weights -> random#50
Open
charleshong3 wants to merge 1 commit into
Open
Conversation
…andom 18k-50k KernelBench-derived baselines initialized their weight/bias tensors to jnp.zeros in create_inputs. With zero weights, `x @ W (+ b)` is identically zero and independent of the input, so the reference output is a trivial constant (all-zero, or a fixed activation thereof). Any kernel returning that constant -- including a no-op that skips the matmul/conv entirely -- passes np.allclose, so these benchmarks could report large meaningless speedups without computing the operator. This replaces the zero-init weights/biases with small-normal random values (~0.02 scale: input-dependent, bf16-representable, no overflow). Only create_inputs is changed; the workload/op is untouched. After the fix a no-op (all-zero output) fails correctness on all 19. Scope: 19 of the affected baselines are fully fixed by non-zero weights. Five others whose *output* is intrinsically small regardless of weights -- the softmax-terminated 38k/43k/50k (row outputs ~1/N) and the structurally degenerate 25k (GroupNorm->Mean) and 42k (Max-Subtract-GELU) -- are NOT addressed here; they need a tolerance or operator change and are left to a follow-up. Megablox (11p) has a distinct input-underflow variant fixed separately. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA). View this failed invocation of the CLA check for more information. For the most up to date status, view the checks section at the bottom of the pull request. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
18k-50k KernelBench-derived baselines initialized their weight/bias tensors to jnp.zeros in create_inputs. With zero weights,
x @ W (+ b)is identically zero and independent of the input, so the reference output is a trivial constant (all-zero, or a fixed activation thereof). Any kernel returning that constant -- including a no-op that skips the matmul/conv entirely -- passes np.allclose, so these benchmarks could report large meaningless speedups without computing the operator.This replaces the zero-init weights/biases with small-normal random values (~0.02 scale: input-dependent, bf16-representable, no overflow). Only create_inputs is changed; the workload/op is untouched. After the fix a no-op (all-zero output) fails correctness on all 19.
Scope: 19 of the affected baselines are fully fixed by non-zero weights. Five others whose output is intrinsically small regardless of weights -- the softmax-terminated 38k/43k/50k (row outputs ~1/N) and the structurally degenerate 25k (GroupNorm->Mean) and 42k (Max-Subtract-GELU) -- are NOT addressed here; they need a tolerance or operator change and are left to a follow-up. Megablox (11p) has a distinct input-underflow variant fixed separately.