[jiaran-king]

Purpose

Enable native decode graphs for Async CAM DeepSeek-V4 target Attention and
DSpark draft Attention. Reuse the vLLM/vLLM-Ascend graph managers and install
AFD metadata at capture/replay boundaries for remote expert calls.

Issue

Related to #499; complements the F-side FULL support merged in #532.

Scope

• Target native graphs with an eager or native-graph DSpark draft; matching
  target/draft TP and independent F execution settings.
• Preserve Attention-local FlashComm1/SP and native PD. P uses the upstream
  complete draft model and existing KV transfer; no new PD protocol is added.
• Accept native FULL/FULL_DECODE_ONLY resolution. Remove graph-specific
  restrictions on compilation mode and model-free speculation; existing
  model-based draft/MTP checks remain. Device qualification uses mode 0.
• Keep DBO excluded and draft graph dependent on target graph, as native target
  capture owns draft capture. Unequal target/draft TP, native kernels,
  quantization changes and lifecycle work remain outside this PR.

Implementation Notes

• Stage target metadata after native capture input preparation and before
  replay, preserving real and padded token counts. Send synchronous DP control
  only when a connector has a control plane; Async CAM has none.
• Scope draft metadata to its own descriptor and query width. Reuse the existing
  physical KV block-size adapter and restore temporary methods/context on
  normal and exceptional exits.
• Native managers retain ownership of graph buffers, buckets, attention/KV
  updates, sampling and fallback. No new ModelRunner, F executor or
  communication protocol is introduced. Preserve main's fix(npu): keep profile routing decisions inside an opaque op #500 routing fix.
• Integrate on main e58fdec5876782b9c6a03683416611deea42af92, retaining its
  MegaMoE path and FFN two-stream removal. No F-runner, connector or
  native-operator changes are included in the PR diff.

Test Plan / Test Result

The conflict resolution at fb4095d merges main 26d8151 and preserves
both Async CAM Attention graph validation and the synchronous CAMP2p
Attention-side routing constraints from #503. Focused CPU checks passed
123 cases, with 35 dependency-gated skips. Changed-file pre-commit
hooks and git diff --check passed. NPU results below are prior evidence;
no new device run was performed for this merge.

After the earlier main synchronization at 7fca98e, focused Async CAM connector/session
CPU checks pass 40 cases; 3 graph test modules are skipped because the
host lacks native dependencies. Documentation hooks, local links and diff
hygiene pass. Conflict resolutions retain the Graph capability descriptions
and main's documentation reorganization. Attention graph implementation files
are unchanged.

The earlier 112 passed, 33 dependency-gated skips and Ruff results at
822779f are reused. The merge retains the graph control-plane test fields and
main's attention_step connector context.

Six pinned-runtime feature-validation checks and the earlier metadata,
restoration, signature and device results below are reused pre-merge evidence.
No new NPU experiment was run for conflict resolution. Configuration tests
continue to admit PD/SP/native options and retain the model-based draft
boundary and DBO rejection.

Retained NPU comparisons used the pre-merge main-integrated execution code
(5afe9dd89314a36850cbb19e9a4e40d9706c3b83, with the admission gates removed):

Combination	Fixed functional requests	Greedy output vs same-config eager	Successful target / draft replays per A rank	Padded draft replays per A rank
SP / FlashComm1	13	All token-for-token identical	73 / 86	86
Native PD + SP / FlashComm1	10	All token-for-token identical	82 / 83	83

Actual SP flags were checked on both target and draft. PD used
MooncakeHybridConnector with native full-draft P TP4. Cases cover one-token,
127/128/129-token and chunked 513-token prompts, concurrency and request reuse.
PD eager/graph both recorded 405 drafted and 263 accepted tokens. In the SP
comparison, the totals were 475/347 and 465/345 respectively; outputs still
matched exactly, and concurrent scheduling was not fixed between arms.

The pinned native configuration resolver was also executed directly:
requested FULL resolves to FULL_DECODE_ONLY for DSV4's UNIFORM_BATCH
backend. This was a configuration check, not another model experiment.
Compilation mode 3 and model-free speculation admission were checked at the
configuration seam only; their device execution is not qualified here.

One formal measurement per arm used W4A8, A DP1/TP4, F EP4/TP1 FULL, five
speculative tokens, 32 requests, 128 input and 128 output tokens, concurrency 4
and four client warmups. Each arm completed 32/32 requests and 4,096 output
tokens. Functional replay probes were removed before timing; their counters
remained unchanged afterward. No profiling ran during measurement.

Combination	Mean TPOT, eager → dual graph	Output throughput, eager → dual graph	Mean TTFT, eager → dual graph
SP / FlashComm1	38.40 → 14.83 ms	86.89 → 221.53 token/s	412.96 → 254.95 ms
Native PD + SP / FlashComm1	36.15 → 12.22 ms	83.85 → 226.11 token/s	1041.01 → 531.84 ms

Retained earlier non-PD, SP-off evidence: nine finite requests matched the
reference token-for-token, with 51 target and 60 draft replays per A rank.
Its single performance comparison measured TPOT 32.20 → 14.63 ms, throughput
102.62 → 226.70 token/s and TTFT 393.14 → 234.11 ms. That earlier measurement
used 140369b873c34779d0d53a8af82aced3c47484eb on an older main baseline;
it was reused rather than rerun. Do not compare absolute numbers across the
separate deployment experiments or attribute the combined gain to one graph.

Pinned stack: vLLM 5df63a7203cdb9328d824a27b682e77c7b9928be, vLLM-Ascend
bd42eea01d8312c4df469fabaed45f77a12df4d3, CANN 9.2 and torch-npu
2.10.0.post7 on Ascend 910C/A3. Existing native binaries were reused unchanged.

MegaMoE with A graphs, W8A8 A graphs, other KV connectors/topologies, GPU
execution, DBO and task-level model accuracy are not newly qualified. Successful cleanup of owned processes
is not a new graceful-shutdown qualification.

Docs Impact

Update the existing Attention, model/platform, Async CAM/FFN and DSpark recipe
contracts. Distinguish actual implementation dependencies from device validation
coverage. Experiment scripts, probes, raw logs and private evidence stay outside
the patch.

---

Essential PR Checklist

• Purpose is clear and linked to public context.
• Scope is bounded.
• Compatibility with the pinned vLLM 0.32 / Ascend pair is considered.
• No upstream source checkouts are changed.
• Existing plugin classes and native graph managers are reused.
• Scoped native hooks are restored, documented and signature-checked.
• Runtime imports retain NPU boundaries and dependency guards.
• Executed, reused and unverified validation paths are distinguished.
• Documentation impact is stated.

[jiangkuaixue123]

Could you clarify the support plan for combining the new target/draft Attention graphs with PD, SP/FlashComm1, and DBO? For each of these, is the current exclusion due to a known implementation limitation, or is it mainly outside this PR's validation scope? What additional work would be needed to enable support, and are there follow-up issues or PRs tracking it?

[jiaran-king]

Could you clarify the support plan for combining the new target/draft Attention graphs with PD, SP/FlashComm1, and DBO? For each of these, is the current exclusion due to a known implementation limitation, or is it mainly outside this PR's validation scope? What additional work would be needed to enable support, and are there follow-up issues or PRs tracking it?

Updated the PR:

• PD/SP: removed the gates after NPU validation. SP (13 requests) and PD+SP (10) matched eager output token-for-token, with target/draft replay and padding confirmed.
• DBO: remains unsupported with DSpark in the pinned runtime and stays blocked.
• Removed unnecessary graph-specific restrictions; native validation remains.

[jiangkuaixue123]

LGTM