docs(RESULTS): speculative wall-clock + decode-framing parity axis - #212
Merged
Merged
Conversation
Two entries: the CPU wall-clock benchmark (the first implementation was O(T^2) and 5.7x slower; int4 kernels lose to int8 on arm64) and the framing finding - dynamic-int8 activation scales are computed per fed tensor, so single-step vs batched decode framings produce materially different decodes on the same machine. Runtime acceptance 0.4614 vs the probe's uniform-framing 0.9886; the probe number is framing-relative. The speculative tier is not viable on quantized CPU artifacts; the playground tier was pulled.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Records the two speculative-decode measurement entries from today's KV-carrying rewrite (interscript-ts#78):
Wall-clock (arm64 CPU, warm, 5 golden rows): plain 2.1 int8 21.4s, lite int4 75.7s (int4 MatMulNBits kernels lose to int8 on CPU - the small drafter is the expensive model here), speculative 155.4s. The first implementation was O(T^2) and 5.7x slower than plain decode despite high token agreement.
Decode framing is a third parity axis. Dynamic-int8 ONNX graphs compute activation quantization scales per fed tensor, so single-step and batched decode framings produce materially different decodes on the same machine, same runtime, same artifact: present values differ ~0.03/element, self-acceptance (drafter==verifier) 33/66, and the int4->int8 pair's batched-verifier output lost a word. Runtime acceptance 0.4614 vs the probe's 0.9886 - the probe figure was measured under uniform plain-path framing and is framing-relative.
Verdict recorded: the speculative tier is not viable on quantized CPU artifacts; the playground tier was pulled (interscript.github.io#185); the runtime keeps SpeculativeModel as measurement infrastructure with the constraint documented.