Add: compose DeepSeek V4 HCA decode TP output

[wangqin1723-max]

• Chain rank-local HCA heads through TP token/head exchange, sharded
  output projection, and dependency-ordered FP32 reduce-scatter.
• Select TP-derived shapes from --tp 1/2/4 at import time and default
  standalone and CI execution to TP2.
• Consolidate the TP output path and fixtures into decode_hca.py,
  reuse the shared output projection, and validate dynamic caches,
  compressed slots, inverse RoPE, and inactive capacity tails.

[coderabbitai]

📝 Walkthrough

Walkthrough

Added a complete DeepSeek-V4 HCA decode output path for context parallelism. The module includes distributed JIT kernels, deterministic fixtures, a PyTorch golden model, active-row and sentinel-tail comparison, and a configurable CLI runner.

Changes

HCA output context parallelism

Layer / File(s)	Summary
Fixtures and validation contracts models/deepseek_v4_flash_dspark/decode_hca_output_cp.py	Defines fixture constants, dimension checks, deterministic tensor specifications, cache data, projection data, and poisoned output buffers.
Distributed HCA output execution models/deepseek_v4_flash_dspark/decode_hca_output_cp.py	Adds device and host JIT paths for sparse HCA attention, attention all-to-all, sharded output projection, and reduce-scatter across physical SP ranks.
Golden comparison and test execution models/deepseek_v4_flash_dspark/decode_hca_output_cp.py	Adds the PyTorch golden model, valid-row comparison, inactive-tail preservation checks, and CLI support for capacity cases and compilation options.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant CLI as CLI test runner
  participant Launcher as l3_decode_hca_output_cp
  participant Kernel as decode_hca_output_cp
  participant Runtime as Distributed runtime
  participant Golden as golden_decode_hca_output_cp
  CLI->>Launcher: Run selected fixture
  Launcher->>Runtime: Allocate communication windows and signals
  Launcher->>Kernel: Execute across SP ranks
  Kernel->>Runtime: Exchange attention data and reduce-scatter output
  CLI->>Golden: Compute reference output
  CLI->>Golden: Compare active rows and sentinel tail

Loading

Possibly related PRs

• hw-native-sys/pypto-lib#801: Implements inverse-RoPE logic in the related HCA output decode path.
• hw-native-sys/pypto-lib#931: Defines related HCA head, projection, and reduce-scatter seams.
• hw-native-sys/pypto-lib#937: Adds a parallel DeepSeek-V4 decode output pipeline for the SWA attention path.

Suggested labels: enhancement

Poem

A rabbit hops through sparse-head light,
While tensors gather rank by rank.
RoPE turns back, projections land,
Tails stay poisoned, neat and dank.
The golden check says, “All is right!”

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly identifies the main change: composing DeepSeek V4 HCA decode output for tensor parallelism.
Description check	✅ Passed	The description directly explains the tensor-parallel HCA output composition and its validation coverage.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@models/deepseek_v4_flash_dspark/decode_hca_output_cp.py`:
- Around line 60-78: Add an import-time ValueError guard alongside the existing
fixture validation checks to ensure CMP_TOPK is greater than
max(FIXTURE_CMP_SLOTS). Use a descriptive message explaining that the fixture
compressed slots exceed the configured top-k capacity, before any fixture
execution or golden validation occurs.

🪄 Autofix

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 9b40ae86-e5d6-4e12-b2fc-3a575a215b1c

📥 Commits

Reviewing files that changed from the base of the PR and between 6eb3b0f and 70d61b9.

📒 Files selected for processing (1)

• models/deepseek_v4_flash_dspark/decode_hca_output_cp.py