GLM-5.3-Flash cookbook: default Blackwell recipes to FP8 KV + TRT-LLM DSA - #36519
Merged
ShangmingCai merged 7 commits intoAug 26, 2026
Merged
Conversation
… DSA Measured on final weights at the rc2 cut: faster on every row (LL +3.5%, HT +2.3-5.5%), 1.8x KV token capacity at identical pool bytes, GSM8K within noise (97.35 vs 97.50). The eight Blackwell cells carry the fp8 / trtllm flags; H100 and H200 stay on BF16 + TileLang, where the fp8 option is disabled and the panel reseats the default onto BF16. The gb300 cells report Verified for either pairing since both were measured.
JustinTong0323
requested review from
sogalin,
wisclmy0611 and
zijiexia
as code owners
August 26, 2026 14:55
github-actions
Bot
added
the
documentation
zijiexia
approved these changes
Aug 26, 2026
- HT FP8 card: add the measured accuracy (GSM8K 97.35%, 99.92% stop) and drop the stale not-re-run claim. - HT BF16 notes: disclose the full extra flag set the measured servers used (the conc 64/256 rows need them; the cell's automatic sizing is not enough). - Precision row: FP8 KV default on Blackwell, BF16 KV on H100/H200. - Badge legend: cover the in-progress state; refresh the FP8 pairing hint. - Reproduce modal summary includes the KV/DSA pairing so the BF16 and FP8 cards read differently.
…ery cell Without it (shared-experts fusion on) the model degenerates on the final weights: empty content, verbatim loops, never stops — isolated on 4x GB300 with the fusion flag as the only difference. Every measured serve behind the published numbers ran with the flag; the minimal-flags trim dropped it by mistake.
Both pairings measured with a 32 GB host tier on the final weights: BF16 + TileLang 1,125.45 / 2,403.57 / 3,839.57 and FP8 + TRT-LLM 1,187.13 / 2,464.89 / 3,941.37 output tok/s at concurrency 16 / 64 / 256 — at most 1.2% below the non-HiCache rows. The random workload never reuses prefixes, so these numbers bound the L2 overhead, not its benefit.
…n flag Both Low Latency rows now come from servers running --disable-shared-experts-fusion, matching the published cells: BF16 1,812.86 tok/s (-0.5% vs the fusion-on measurement) and FP8 1,870.82 tok/s (-0.8%); TPOT unchanged, FP8 stays 3.2% ahead. The flag costs nothing measurable on throughput.
ShangmingCai
approved these changes
Aug 26, 2026
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
3 participants
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.
Motivation
Follow-up to #36513 (merged) — the default-switch commit missed that squash. Since then three more issues surfaced in the published cookbook, so this PR now carries the full corrective batch:
--disable-shared-experts-fusion— with fusion ON the model silently degenerates on the final weights (empty content, repetition loops, never stops). Isolation on 4x GB300: identical serves, only this flag differs — ON degenerates, OFF answers and stops. Every measured serve behind the published numbers ran with the flag.AttributeError: DSATokenToKVPool has no attribute full_kv_pool) — the HiCache options are disabled for Low Latency until the code fix lands (validation in flight). High Throughput keeps them.Modifications
--kv-cache-dtype fp8_e4m3+ TRT-LLM DSA; H100/H200 keep BF16 + TileLang (FP8 option disabled on Hopper, reseats automatically).isRecommendedSelectionis hardware-conditional; GB300 shows verified for both pairings.--disable-shared-experts-fusionto all twelve cells.Tests
node docs/scripts/check_cookbook_configs.mjs— OK;mint validate— passed.CI States
Latest PR Test (Base): ⏳ Run #32991107873
Latest PR Test (Extra): ⏳ Run #32991107656
Latest PR Test (AMD ROCm 7.2): ➖ No AMD PR run found for this commit.