Skip to content

[model, ckpt] feat: Add DeepSeek V4 Bridge support - #2

Closed
bobboli wants to merge 8 commits into
bridgefrom
fix/dsv4-r3-config
Closed

[model, ckpt] feat: Add DeepSeek V4 Bridge support#2
bobboli wants to merge 8 commits into
bridgefrom
fix/dsv4-r3-config

Conversation

@bobboli

@bobboli bobboli commented Aug 19, 2026

Copy link
Copy Markdown
Owner

Summary

  • add on-the-fly dequantization utilities for Hugging Face checkpoints
  • add the DeepSeek V4 Bridge, conversion examples, documentation, and tests
  • include the DeepSeek V4 RoPE configuration and pipeline-parallel mapping fixes
  • map the hash-routing layer count to the Megatron-Core dsv4_n_hash_layers field

Scope

This draft targets the latest bridge branch to expose the complete dependency diff for review. It contains eight commits across the DeepSeek V4 support and follow-up fixes.

Verification

  • uv run python -m pytest tests/unit_tests/models/deepseek/test_deepseek_v4_bridge.py -q: 16 passed
  • targeted pre-commit hooks: passed

cuichenx and others added 8 commits August 17, 2026 17:54
@cuichenx @bobboli
[model, ckpt] feat: Add on-the-fly dequantization utilities for HF ch…
d557d10 
…eckpoints (NVIDIA-NeMo#3778)

Signed-off-by: Chen Cui <chcui@nvidia.com>
@weijiac0619 @claude
@cuichenx @bobboli
DeepSeek V4 Bridge (NVIDIA-NeMo#3562)
2330368 
Signed-off-by: weijiac <weijiac@NVIDIA.com>
Signed-off-by: weijiac <weijiac@nvidia.com>
Signed-off-by: chcui <chcui@nvidia.com>
Signed-off-by: Chen Cui <chcui@nvidia.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: chcui <chcui@nvidia.com>
@Meirtz @claude @bobboli
fix(model): keep full rotary_percent for DeepSeek-V4 MLA rope (NVIDIA…
30ab915 
…-NeMo#4271)

Signed-off-by: Lingrui Mei <lmei@nvidia.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
@HollowMan6 @bobboli
DSV4: Making compress_rope_theta priotitzed than rope_scaling->rope_t…
3be1707 
…heta (NVIDIA-NeMo#4802)

Signed-off-by: Hollow Man <hollowman@opensuse.org>
@bobboli
fix(deepseek-v4): support nested rope theta config
2cd1845 
Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com>
@bobboli
[conversion] fix: skip absent auto mappings across PP
ca1a03d 
Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com>
@bobboli
[model] fix: Map DeepSeek V4 hash layer config
6f48f43 
Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com>
@bobboli
test: clarify isolated DSV4 export coverage
8eaf211 
Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com>
@bobboli

bobboli commented Aug 19, 2026

Copy link
Copy Markdown
Owner Author

Superseded by the clean backport in #5 and the focused hash-layer follow-up in #3. The branch is retained for reference.

@bobboli bobboli closed this Aug 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

5 participants

@bobboli @cuichenx @weijiac0619 @Meirtz @HollowMan6