Repository navigation
feat: add DeepSeek V4.1 B300 TP2/TP4 operator profiles - #233
Merged
Merged
Conversation
|
Note Currently processing new changes in this PR. This may take a few minutes, please wait... ⚙️ Run configurationConfiguration used: Repository: ai-dynamo/aisimulate/.coderabbit.yaml Review profile: ASSERTIVE Plan: Enterprise Run ID: ⛔ Files ignored due to path filters (16)
📒 Files selected for processing (28)
Comment |
This was referenced Sep 16, 2026
Harrilee
marked this pull request as ready for review
September 17, 2026 16:13
simone-chen
approved these changes
Sep 18, 2026
Harrilee
force-pushed
the
feat/harrli/dsv41-silicon
branch
from
September 18, 2026 21:48
5b315ab to
b7d65dc
Compare
Harrilee
force-pushed
the
codex/dsv41-b300-op-tp2-tp4
branch
2 times, most recently
from
September 18, 2026 22:56
0cad95e to
472b4cb
Compare
Harrilee
force-pushed
the
feat/harrli/dsv41-silicon
branch
2 times, most recently
from
September 21, 2026 02:59
ad7ffee to
ae8905b
Compare
Harrilee
force-pushed
the
feat/harrli/dsv41-silicon
branch
from
September 21, 2026 03:10
ae8905b to
d7869f4
Compare
Harrilee
changed the title
Harrilee
force-pushed
the
codex/dsv41-b300-op-tp2-tp4
branch
from
September 21, 2026 04:18
472b4cb to
4206388
Compare
Contributor
Author
|
Rebased the B300 supplement onto latest OP parent #160 ( Validation on this head: 120 Collector tests and 386 native parity cases passed. All 6,444 stored module/GEMM/MoE/NCCL rows reproduce exact timings through strict native SILICON queries, with sidecar hashes and physical-key uniqueness checked. Both Fast CI runs passed; Full CI is running, not yet a pass. There were no actionable review comments or unresolved inline threads. Updated the PR title to the repository's |
Harrilee
enabled auto-merge (squash)
September 21, 2026 04:34
Harrilee
force-pushed
the
codex/dsv41-b300-op-tp2-tp4
branch
from
September 21, 2026 04:36
4206388 to
75ad157
Compare
Contributor
Author
|
Resolved the conflict caused by #160's squash merge and this PR's retargeting to This update changes ancestry only. Its complete Git tree ( The body now reflects the merged prerequisite and direct main base. New-head Full CI is running; previous-head success is not represented as passing the new commit's checks. No new accuracy qualification is claimed. |
This was referenced Sep 21, 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
2 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.
Summary
Adds measured B300 TP2/TP4 operator databases for DeepSeek-V4.1-Flash SILICON prediction, with separate full and bounded execution profiles. The OP prerequisite #160 is merged; this PR now targets
mainand retains the existing GB300 data. The Collector derives GEMM/MoE/NCCL keys from the loaded TP layout and rejects mismatched process groups, expert topology and vocabulary padding before measurement.Current head:
75ad15780b4a163070f9a75b6aae3bc904b8dbbf. Rebased the B300 supplement ontomain(3bfdbf88) after #160 was squash-merged. This only updates ancestry: the complete Git tree is identical to previously validated42063883(ed777e702998f78ac7baac723e741e2fdcbd51a7). All code, tests and 58 profile files, including 24 Parquets, are unchanged. The PR contains no raw campaign artifacts or end-to-end ground truth.Validation — 2026-09-21
The local results below were obtained at
42063883, whose entire source/data tree is identical to the current head. The ancestry-only rebase did not rerun those tests or GPU collection.source='silicon'. This covers all six profiles and all 24 tables: 5,824 module, 248 GEMM, 124 MoE and 248 NCCL rows. Sidecar hashes, row counts and unique physical keys also pass.42063883: both Fast CI runs and Full CI passed. Both new-head Fast CI runs passed (PR run, trusted-copy run). New-head Full CI remains pending; previous CI does not satisfy the new commit's gates.There are no unresolved inline review threads or actionable review comments. The existing approval predates this rebase. The PR remains ready for review; #160 is already merged.
Exact table queries are integrity checks, not holdout accuracy. This rebase performs no GPU recollection or new MAPE evaluation. Historical CI and wheel evidence do not substitute for validation of this head; the results below retain their original sources, populations and limitations.
Packaged data
python/aisimulate/src/aisimulate_core/systems/profiles/dsv41/tp2/{full,decoder_bounded}; TP4 roots:python/aisimulate/src/aisimulate_core/systems/profiles/dsv41/{full,decoder_bounded}. Select the documented B300 system, TP, runtime version and decoder profile explicitly.Historical packaging validation reproduced all 2,244 TP2 records and all 322 full / 326 bounded qualified heldout predictions from a built wheel. That is retained collection evidence, not a claim that wheel validation was repeated by this rebase.
Historical native-forward validation
All methods use the same independently collected heldout observations within each TP/profile: 38 predeclared geometries, ten fixed attempts each. MAPE is the unweighted mean of geometry errors against the median of all qualified fixed attempts. Failed native geometry/batch-split attempts remain in the coverage denominator; they are excluded from numerical error, never counted as zero error. These are native-forward latencies, not HTTP serving metrics or a blind benchmark.
FPM aggregate coverage is separate. Full aggregate and static results agree. Bounded aggregate retains its original multiple-prefill guard: TP2 MAPE 6.7164%, 26/38 shapes and 260/380 attempts (66 otherwise-qualified guard rejections plus 54 native failures); TP4 MAPE 1.6108%, 26/38 and 260/380 (86 guard rejections plus 34 native failures). Static predictions use actual per-request seed/chunk/context witnesses; no aggregate guard was bypassed.
FPM calibration and heldout have disjoint run IDs, request IDs and exact geometries: 145 calibration versus 38 heldout shapes. Rows use the predeclared first qualified fixed attempt, never the fastest. Calibration qualified 1,224/1,450 TP2 full, 1,248/1,450 TP2 bounded, 1,275/1,450 TP4 full and 1,280/1,450 TP4 bounded attempts. All failures remain retained. No heldout latency is a predictor input, no heldout fitting occurred, and hardware/TP data were not substituted across cells.
The TP2 bounded FPM worst-point error is 138.5442% at decode B2 / total past-KV 3,584: prediction 272.4125 ms versus heldout median 114.1979 ms. This is the native interpolation between the first-qualified calibration observations at 3,072 (433.5766 ms) and 4,096 (111.2484 ms). Both original trace values were verified; later faster samples were not substituted and the slow sample's cause is not established. TP4 full retains its previously disclosed 111.33% worst-point error.
Measurement identity and evidence
B300 is
b300_sxm, separate from GB300. TP4 usesdev-c4ca651192e5; TP2 usesdev-c4ca651192e5-tp267c0e788b4b2, binding the same immutable image to its separately qualified startup source/policy. The complete image/runtime/source identities, original package-reported version, driver, raw-admission and prediction hashes are in compact sidecars. Image hashes identify builds, not an asserted upstream Git revision.TP2 OP jobs
4363517(full) and4364038(bounded), and FPM jobs4363666(full) and4364952(bounded), completed with exit 0. Each FPM archive contains 11,180 files; every remote hash was recomputed locally. Both roles passed original source, token-freeze, trace, ten-attempt, physical-pool, prerequisite and identity gates. Raw-copy negative controls separately rejected invalid timing, altered tokens, pool size, prerequisite handoff and campaign role, including rebound outer hashes. Source-bound adapter corrections preserve the original producer and measured bytes.The historical TP2 predictions used SOL
40fdb603, SILICON6ad6af2d, and FPMf2aea3a1/76c5a5e2. Independent review at collection time recomputed the observed medians, attempts, MAPE, selection and worst-point trace evidence. TP4 observations and recorded prediction values are retained. These source identities remain the scope of the accuracy evidence; current-head rebase validation is reported separately above. OP collective baselines record Torch NCCL 2.29.7; actual FPM workers independently record SGLang PyNCCL 2.30.7. Their distinct library identities are preserved.Companion FPM data PR: #232, targeting #158. This OP PR does not add GB300 TP2 measurements.