bf16 PPGD source: quality A/B + full32L memory cash-in configs (lp-bf16-source-ab)#938
Open
ocg-goodfire wants to merge 4 commits into
Open
bf16 PPGD source: quality A/B + full32L memory cash-in configs (lp-bf16-source-ab)#938ocg-goodfire wants to merge 4 commits into
ocg-goodfire wants to merge 4 commits into
Conversation
…PGD-bsc, 10k steps) fp32 control vs bf16 treatment (source_dtype knob a3ffa2a), same seed, dp8, slow_every 2000 for a PGD_20step adversary-effectiveness trajectory. Bridge task: lp-bf16-source-ab. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…attempt) mem_profile-and-exit variants of llama8b_full32L_HSDP_b128_dp32 with source_dtype bfloat16. Bridge task: lp-bf16-source-ab. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Experiment configs for the bf16 PPGD source+moments investigation (bridge task
lp-bf16-source-ab, items #1+#2 of the lower-precision roadmap). No code changes — thesource_dtypeknob (a3ffa2a) is exercised as-is.What's in here
pile_ppgd_bsc_srcdtype_ab_{fp32,bf16}.yaml— Tier-1 quality A/B on the pile-4L PPGD-bsc lineage (10k steps, dp8, same seed; only diff =source_dtype: bfloat16)llama8b_full32L_HSDP_{b128,b160}_dp32_bf16src_MEMPROBE.yaml— mem_profile-and-exit probesllama8b_full32L_HSDP_b160_dp32_bf16src_TPUT.yaml— 60-step b160 throughput probeResults
Quality gate PASSED (runs p-1b097fd2 fp32 / p-07ae1175 bf16): bf16 arm equal-or-better on every judge metric — persistent-adversary recon within ±2% from 5k steps, fresh-PGD-20step eval −1..−13%, masked CE/KL −2..−8%, CI-L0 ~8% sparser, faithfulness identical. Post-hoc ckpt probe: NO Adam v underflow (bf16 shares fp32's 8-bit exponent; the risk was mantissa swallowing, which didn't materialize — v/m stats match fp32).
Memory confirmed (p-910a7582): b128/dp32 resident 49.5 → 35.25 GiB (−14.25 GiB exactly), runtime peak 148.2 → 132.2 GiB.
b160 / 5 seq/GPU now FITS (p-03721a2f): runtime peak 144.8 / 164.1 GiB (fp32 sources OOM'd at ~172). Throughput probe p-8dd01149 in flight.
Full trail: bridge task log
lp-bf16-source-ab.🤖 Generated with Claude Code