Kaggle T4x2 Mistral RULER multi-needle paired beneficial-quant kernel (4K, n>=50)#25
Draft
jagmarques wants to merge 2 commits into
Draft
Kaggle T4x2 Mistral RULER multi-needle paired beneficial-quant kernel (4K, n>=50)#25jagmarques wants to merge 2 commits into
jagmarques wants to merge 2 commits into
Conversation
… (4K, n>=50) Adds experiments/kaggle/nq_mistral_ruler_niah/: a RULER-style multi-needle NIAH kernel for Mistral-7B-Instruct-v0.3 at 4K context with n=50 trials and 8 needles per trial. Tests FP16, K3V2_pb0, and K4V2_pb0 via a paired protocol (single prefill, clone+quantize per config) to produce defensible beneficial-quant evidence. Results pending queued Kaggle run.
8-needle variant hit 48/50 FP16 (saturated). Bumping to 24 needles increases retrieval interference in the same 4K context to push FP16 below saturation and expose beneficial-quant deltas.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
RULER-style multi-needle paired beneficial-quant kernels for Mistral-7B-Instruct-v0.3 at 4K (8-needle and 24-needle). Real E8 quant (inverse_rope -> Hadamard -> per-head amax -> E8Lattice.nearest_point -> forward_rope), paired protocol (single prefill, clone+quantize KV per config), n=50, value-specific recall.
RESULTS (both COMPLETE, re-derived from cells, errors={}, anti-vacuous clean):
HONEST FRAMING: these are LOSSLESS confirmations on a multi-needle 4K task, NOT a beneficial-quant rescue result. FP16 saturates at 96-98% at 4K regardless of needle count, so there is no FP16 dead zone to rescue and paired McNemar is non-discriminative (b approx c). K3V2 50/50 at 24 needles is within noise, not a claimable win.
CONCLUSION: the rescue-bulletproofing test needs an FP16 dead zone, which requires long context (8K-32K) that OOMs on T4x2 at depth. It is A100-bound. Kept DRAFT: the harness is reusable for that A100 run; no headline claim, ties to no criterion, so not merged.