Repository navigation
[Fix] GLM-5.3-Flash F4.1: Correct mHC precision and replica gradient reduction - #2156
Open
bingshao333 wants to merge 5 commits into
Open
bingshao333 wants to merge 5 commits into
bingshao333 wants to merge 5 commits into
Conversation
Adds xtuner/v1/module/decoder_layer/mhc.py (MHCConfig / hc_split_sinkhorn / hc_pre / hc_post / unshard_hc_params) and xtuner/v1/ops/hc_post.py (fused Triton hc_post_fused), following doc/xtuner_glm5p3flash_design.md F4. The core math is ported from xtuner's dsv4 branch (DeepSeek-V4's Hyper- Connections, commit 01c31a8, not merged into this branch) into the model-agnostic public location the design doc calls for, dropping V4's XTUNER_V4_HF_PARITY global toggle (the default bf16-fast path already degrades to HF-exact math under fp32 inputs, which the tests use as the parity anchor) and the unported 721-line TileKernels backend (left as a documented NotImplementedError gap rather than a blind port with no hardware to validate it against). Adds xtuner/v1/model/moe/glm53/decoder_layer.py: Glm53DenseDecoderLayer (overrides DenseDecoderLayer._forward) and Glm53MoEDecoderLayer, which only overrides MoEDecoderLayer's _pre_moe_forward/_post_moe_forward seams so the ~400-line EP/dispatcher/domino-micro-batch pipeline stays untouched; the mHC residual rides through those methods as an opaque _MHCResidual payload in place of the base class's plain Tensor. mhc_cfg=None degrades both layers to the base class's ordinary residual math unchanged, for reuse by the MTP layer in F6. Also fixes a real bug surfaced while testing: KDA's dt_bias/A_log were left torch.empty-uninitialized in xtuner/v1/module/attention/kda.py; now initialized to match HF's Glm5NextTextForgetGate _init_weights policy (A_log zeroed when a safe gate lower bound is set, dt_bias log-uniform). Verified with 26 tests across tests/model/test_glm53_mhc.py (Sinkhorn math + exact HF parity in fp32, bf16 parity within ULP tolerance), tests/ops/test_hc_post.py (fused kernel vs eager reference, forward/ backward/compile), and tests/model/test_glm53_decoder_layer.py (both decoder layers, mhc_cfg=None passthrough equivalence, gradient flow). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
HF normalizes the flattened mHC streams with the text-config RMS epsilon, not the hyper-connection eps. The hc_pre norm_eps parameter existed but all four call sites relied on the 1e-6 default, so checkpoint gradients drifted from HF. Pass input/post-attention layernorm variance_epsilon at each site; the five-layer crop gradient oracle now agrees with HF within 5%.
Select the F4 mHC primitive and regression-test changes from JT-Ushio/xtuner-muonsplit-verify commit a8d798a. Preserve the original author and add type annotations to the new test.
(cherry picked from commit 46e207b)
jayhenry
force-pushed
the
feat/glm53flash-f4-mhc
branch
from
October 10, 2026 19:01
6841448 to
b245bb8
Compare
This branch has not been deployed
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
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.
This F4.1 follow-up aligns mHC numerical behavior with Hugging Face and fixes gradient synchronization for FSDP-ignored non-expert parameters when EP > 1 or expert TP (ETP) > 1.
Target:
feat/glm53flash-f4-mhc. Related: #2107 and #2136.Changes and impact
hc_fnprojection in FP32, avoiding the precision loss from BF16 projection, and return the collapsed stream in the original activation dtype. Add HF forward/backward comparisons with FP32 and BF16 projection weights, plus configured RMS normalization epsilon coverage at all four Dense/MoE attention/FFN call sites. The epsilon wiring is already fixed in the target branch; this PR adds its regression coverage.Replicatesubmesh that omits their FSDP replicas. For fully replicated non-expert parameters whose mesh is smaller than the model mesh, average gradients over the complete replica group. This shared MoE fix also covers affected ignored parameters beyond mHC. Add regression coverage for replica-mean gradients and two SGD updates across EP/SP token partitions, preserving ordinary sharded and expert gradient behavior.Validation