perf(qwen36): add SpecLA factor-capture speculative decoding

[Graffioh]

Current gain over main

Qwen3.6-27B Q4_K_M target, Q8_0 DFlash draft, DDTree budget 22, top-k 4, tau=6, ten HumanEval-style prompts, and 128 generated tokens:

Configuration	Mean decode throughput
Current main	25.14 tok/s
SpecLA	27.19 tok/s

SpecLA gain over current main: +8.2%.

Paper: https://arxiv.org/abs/2607.16673

CLI

The normal entry point is one flag:

dflash_server TARGET.gguf --draft DRAFT.gguf --specla

--specla enables speculative linear-attention verification when the loaded target and placement support it. SpecLA is the verification/state-management mode; it is not an alias for a particular proposal algorithm.

For the Qwen3.5/Qwen3.6 dense path implemented in this PR, the runtime currently:

• selects DDTree as the compatible proposal adapter;
• enables fast rollback;
• uses DDTree budget 22;
• uses confidence-pruning margin tau=6;
• uses draft-tree top-k 4.

This separation is intentional. A future Qwen3.8 DSpark implementation can provide candidates and parent topology behind the same --specla flag without changing the user-facing CLI. That DSpark adapter and its target-specific recurrence/factor capture are not part of this PR.

Supported and permitted combinations

The path in this PR is supported for a monolithic Qwen3.5/Qwen3.6 dense target on one local device. It has been built and parity-tested on HIP.

The current Qwen adapter requires:

• --draft DRAFT.gguf;
• fast rollback, which --specla enables automatically;
• a non-split, single-device target placement.

Setup-dependent overrides remain available:

Option	Constraint	Default with --specla	Purpose
--ddtree-budget N	positive integer	22	Maximum DDTree proposal nodes
--ddtree-tau T	finite, non-negative number	6	Cumulative log-probability pruning margin
--specla-top-k K	positive integer	4	Draft-tree expansion width
--draft-swa N	positive window when needed	checkpoint metadata	Override draft sliding-window metadata for older checkpoints

The environment variables DFLASH_SPECLA=1 and DFLASH_SPECLA_TOPK=K remain compatibility controls for non-CLI harnesses. DFLASH_SPECLA_CONDITIONAL_DRAFT=1 and DFLASH_SPECLA_FUSED_COMMIT=0 are advanced experiment/debug controls, not normal launch requirements.

Rejected or unavailable combinations

• --specla --no-fast-rollback is rejected because SpecLA requires the rollback/state path.
• --specla-top-k without --specla is rejected.
• The test_dflash harness rejects --specla --seq-verify; sequential verification cannot exercise the SpecLA tree path.
• The current Qwen adapter rejects a missing --draft.
• Multi-device or layer-split SpecLA is not implemented by this PR.
• --kvflash <tokens|auto> cannot use SpecLA factor migration. The server warns, reports SpecLA off, and falls back to ordinary DDTree verification.
• Paged-attention autoregressive mode cannot be combined with the current draft/DDTree speculative path.
• On an architecture or placement without a SpecLA adapter, the server prints a warning and uses that architecture’s normal decode path. An independently requested --ddtree mode remains independent.

Paper mapping

The paper requires scored speculative candidates and parent topology, not DDTree specifically. A chain is a degenerate tree. DDTree is simply the proposal adapter connected for Qwen3.6 today; DSpark can be connected later if it supplies the required candidate/topology contract and the target has a compatible SpecLA recurrence and factor-capture implementation.

[Graffioh]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@Graffioh I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

1 issue found across 33 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="server/src/qwen35/qwen35_dflash_target.cpp">

<violation number="1" location="server/src/qwen35/qwen35_dflash_target.cpp:609">
P2: The SpecLA bank-rotation / `specla_compact_fused` / `specla_flush_raw_fused` logic is copied nearly verbatim between this file and `server/test/test_dflash.cpp`. `rollback_to_tree` (walked_sibling compaction + `current_bank`/`old_pending_bank` rotation + `pending_count` update), `finish_speculative_state` (bank selection via `specla_pending_bank` + `specla_flush_raw_fused`), and `rollback_to_specla` (bank toggle + count) each have a hand-maintained mirror in the bench harness. These are the two most state-sensitive routines in the PR; any divergence between the production target and the harness (e.g. the `ne` bounds, channel mapping, or bank-toggle logic) silently changes acceptance behavior. Pull the shared commit/flush helpers into a common function (e.g. in `specla_commit_cuda.h/.cu`) and have both call sites invoke it.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[Graffioh]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@Graffioh I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

All reported issues were addressed across 36 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 36 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 37 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>