You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Trainer half of the DSV4 canonical-fold unification (serving half: togethercomputer/xorl-sglang#18). Built on #45 plus a merge of #44, this migrates DSV4 off the architecture-scoped NCCL-tree contributor-order reproduction onto the shared canonical_moe_fold_v1 — the deliberate byte change #45 flagged as requiring full requalification, which this branch carries end to end (K3 exactly 0.0 at the 64-decision trained-adapter promotion replay, k3_max 0.0, decode 5.5 tok/s).
Fold switch (1adc5db7)
exchange_variable_and_nccl_tree_chain_sum ([1..7,0] left-associative chain) is retired; exchange_variable_and_canonical_fold keeps DSV4's variable-row all-to-all transport and reduces the rank-ordered arrivals with canonical_moe_fold_v1 — the same byte program serving evaluates through its gated canonical post-experts combine.
New unit witnesses include a byte-divergence test proving the fold is the balanced tree and not any left-associative chain (half-ulp tie construction).
Launch-geometry mirroring (facace92)
The trainer's Marlin row pad to 48 is retired: under the chunked exact-Marlin program (xorl-sglang#18) the pad made the trainer's decode segment run its live row in an M=10 chunk while serving launches M=1, diverging value-dependently. The trainer now launches exactly the serving runner's M per gathered segment.
Exact-head scoring kernel (41bd2c54)
Serving's deterministic mode interposes log_softmax with the batch-invariant Triton kernel, whose BF16 rounding differs from ATen's on boundary rows (proven at one decision: the f64 truth sits 5.5e-8 past the bf16 boundary; the entire trunk through final-norm was byte-equal). The DSV4 exact head's forward VALUE now uses the serving kernel; the surrogate VJP keeps FP32 reference math.
Qualification record (e940116f, plan doc + lane log)
Base ruler byte-equal at 4 AND 64 decisions (engagement witness: decode bytes changed vs the DSV4-Flash active-LoRA zero-K3: full qualification closed (K3 = 0 at 64-decision promotion) #45 program); A-join zero/nonzero joins green with the perturbed negative control correctly diverging; training gate with a correctly-diverging post-step control; B-join with the saved dsv4_expert_banks adapter; 64-decision promotion at K3 = 0.0.
Three latent campaign-1 value-luck defects were burned down along the way (off-path serving combine gate, Marlin multi-block-expert completion-order race, ATen-vs-BI log_softmax split); the lane log carries the full localization record.
Merge conflicts with #44 resolved per the reconciliation notes (auto.py keeps both the Qwen capability resolution and the DSV4 fail-closed registration; single-writer + 948-factor export both preserved). Supersedes the combine-order half of #45; intended for the #41 -> #43 -> #44 -> #45 restack by the reconciliation owner.
We're continuously improving Broly's accuracy and finding quality — your feedback is valuable. False positives, missed findings, bugs, and feature requests all welcome.
Folded into #45 by the reconciliation restack. Final trainer head: b8b73ef4a217d9ce8e423e5d2f430b2fbcffbf57, pinned to xorl-sglang #17 at 9680d4bda91ad21f5eb592d02eeebcbe2d901e4f. The production Campaign 2 summary is retained; campaign-only lane artifacts remain out of the public diff.
Final reconciled heads after the live gitlink and hygiene audits: trainer #45 is d09fb81540bde8e3689c14c725e8229a301e2643, pinned to sampler #17 at 9680d4bda08bb7b9878aec1642d74688db36dfaa. The campaign-only hash-topk replay helper was removed from the public diff to close Broly findings.
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
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.
Trainer half of the DSV4 canonical-fold unification (serving half: togethercomputer/xorl-sglang#18). Built on #45 plus a merge of #44, this migrates DSV4 off the architecture-scoped NCCL-tree contributor-order reproduction onto the shared
canonical_moe_fold_v1— the deliberate byte change #45 flagged as requiring full requalification, which this branch carries end to end (K3 exactly 0.0 at the 64-decision trained-adapter promotion replay, k3_max 0.0, decode 5.5 tok/s).Fold switch (
1adc5db7)exchange_variable_and_nccl_tree_chain_sum([1..7,0] left-associative chain) is retired;exchange_variable_and_canonical_foldkeeps DSV4's variable-row all-to-all transport and reduces the rank-ordered arrivals withcanonical_moe_fold_v1— the same byte program serving evaluates through its gated canonical post-experts combine.Launch-geometry mirroring (
facace92)Exact-head scoring kernel (
41bd2c54)log_softmaxwith the batch-invariant Triton kernel, whose BF16 rounding differs from ATen's on boundary rows (proven at one decision: the f64 truth sits 5.5e-8 past the bf16 boundary; the entire trunk through final-norm was byte-equal). The DSV4 exact head's forward VALUE now uses the serving kernel; the surrogate VJP keeps FP32 reference math.Qualification record (
e940116f, plan doc + lane log)dsv4_expert_banksadapter; 64-decision promotion at K3 = 0.0.Merge conflicts with #44 resolved per the reconciliation notes (auto.py keeps both the Qwen capability resolution and the DSV4 fail-closed registration; single-writer + 948-factor export both preserved). Supersedes the combine-order half of #45; intended for the #41 -> #43 -> #44 -> #45 restack by the reconciliation owner.