WIP: reconcile GLM-5.2 glm_moe_dsa (spec-decode + #130)

[drunkcoding]

DRAFT — do not merge. Base: dev. This PR reconciles the two divergent glm_moe_dsa
implementations (spec-decode's, already on dev via #140; and PR #130's) into one canonical
stack by salvaging only #130's genuinely-unique, non-conflicting deltas onto dev.

Strategy

dev @ 0e4904f (post-#140/#150) is the canonical baseline: it already registers glm_moe_dsa
(type 5), ships its own DSA/MTP-integrated SyncGlmMoeDsaMoEBlock, the native FP8 e4m3 block-scale
dequant-on-copy path, DSA indexer/IndexShare, and MTP spec-decode. PR #130 branched from an older
dev (b33d5fa) and re-implemented an overlapping-but-divergent GLM/FP8 stack. Per the written
reconciliation plan, we canonicalize on dev and salvage only #130's unique wanted deltas —
dropping everything spec-decode already provides.

Full auditable triage (git cherry, range-diff, per-commit verdicts, and every deviation with evidence)
is in docs/superpowers/plans/2026-08-14-glm-5.2-reconciliation.decision-log.md (in this PR).

What landed (net wanted deltas) — 4 files

File	Source	What
moe_infinity/utils/fp8.py	#130 707cc06 (subset)	Append dequant_fp8_state_dict + stack_expert_scales (+re/typing/EXPERT_SCALE_KEY_RE). Byte-identical to #130's tested helpers; dev's dequant_fp8_blockwise untouched.
core/model/model_topology.cpp	#130 b5a015a	Coalesce sparse-expert partition reads into one sequential bulk read per file (struct TensorPlacement + posix_fadvise(SEQUENTIAL)); fixes multi-day cold-cache store reloads. clang-format clean.
tests/python/unit/test_glm_moe_dsa.py	#130 9b49f43	Reconciled unit suite — the only coverage for the two new fp8 helpers. Two guards adapt it to dev (below).
docs/superpowers/plans/2026-08-14-glm-5.2-reconciliation.decision-log.md	new	Evidence + verdicts + deviations.

Per-commit verdicts (9 unique #130 commits)

git cherry -v origin/feat/dflash-spec-decode origin/feat/glm-5.2-fp8-support → 9 + commits:

SHA	Verdict	Why
707cc06 fp8 utils	PARTIAL-KEEP	Keep only dequant_fp8_state_dict + stack_expert_scales; dev owns dequant_fp8_blockwise (d797138).
038c149 register	DROP	dev already registers glmmoedsa (constants.py, hf_config.py).
98a5a39 block	DROP (alt)	dev owns SyncGlmMoeDsaMoEBlock (b02c3bd); #130's is a divergent alternative.
c8ae160 on-GPU dequant	DROP (alt)	Canonical = dev's dequant-on-copy (3167500+c16cb46, Option A==B parity).
b5a015a partition-read perf	KEEP	model_topology.cpp untouched by spec-decode; self-contained.
187048c offload wiring	DROP (alt)	Superseded by dev's 5b2c2b6.
9b49f43 tests+docs	KEEP (reconcile)	Unit suite is unique; README row/note + registry assertion already on dev.
3dc4e72 NVTX	DROP (build-breaking as shipped)	See deviation 3 — vendored header breaks the build; superseded by ba56518.
3444ed9 smoke e2e	DROP (covered)	dev's test_glm_smoke.py already has the MOE_GLM_SMOKE gate.

Divergence summary (spec-decode/dev vs #130)

• Block API: dev's SyncGlmMoeDsaMoEBlock delegates routing to HF (_route → _hf_route_tokens)
  and integrates DSA/MTP; feat(glm): support zai-org/GLM-5.2-FP8 (glm_moe_dsa) with FP8-in-store expert offload #130's inlines route_tokens_to_experts + nvtx annotations and assumes an
  always-present executor. Kept dev's.
• FP8 path: dev keeps routed experts FP8 in-store and dequantizes on-copy in the native
  dispatcher (c16cb46); feat(glm): support zai-org/GLM-5.2-FP8 (glm_moe_dsa) with FP8-in-store expert offload #130 dequantizes on-GPU after fetch in a competing C++ path (touches the
  same expert_dispatcher/expert_module/py_archer_prefetch). Kept dev's; feat(glm): support zai-org/GLM-5.2-FP8 (glm_moe_dsa) with FP8-in-store expert offload #130's C++ dropped.
• DSA attention: dev ships a dedicated glm_dsa.py + DSA indexer/IndexShare; feat(glm): support zai-org/GLM-5.2-FP8 (glm_moe_dsa) with FP8-in-store expert offload #130 has none.

Deviations from the plan (dev evolved past plan assumptions — all evidence-backed)

1. Task 4.2 (README row/note + registry assertion) — SKIPPED, already on dev. README already has
   a GLM-5.2 table row + a more detailed note + a full usage section; test_model_registry.py lines
   29-30 already assert glmmoedsa. Re-adding would duplicate (violates the no-dup rule). README.md
   and test_model_registry.py left untouched.
2. test_glm_routing_parity_vs_hf — GUARDED. It calls sync.route_tokens_to_experts(...) (feat(glm): support zai-org/GLM-5.2-FP8 (glm_moe_dsa) with FP8-in-store expert offload #130's
   block API); dev's canonical block has no such method (hasattr == False) and this env ships
   transformers.glm_moe_dsa, so verbatim would AttributeError. Added a hermetic class-level
   hasattr guard that skips cleanly (parity is already covered by
   test_glm_routing.py::test_routing_parity, which uses block._route).
3. NVTX salvage (3dc4e72) — DROPPED with compile-probe evidence. The partially-vendored
   nvtx3.hpp (only nvtx3.hpp, not its #include "nvToolsExt.h" sibling) shadows the working
   system header (dev already puts core/include first on the include path) and fails to compile
   on this CUDA-13.1 host: P1 (vendored wins) → fatal error: nvToolsExt.h: No such file or directory;
   P2 (dev current) → OK. dev already solves CUDA≥12.9 NVTX header-only via ba56518 (does not link
   libnvToolsExt); feat(glm): support zai-org/GLM-5.2-FP8 (glm_moe_dsa) with FP8-in-store expert offload #130's setup.py hunk would reverse that and re-add fp4-kernel/CUDA-stub/
   python_requires>=3.8 regressions if blind-applied. Landed neither file. Reviewer options in the
   decision log.

Verification

• Unit suite (required): test_glm_moe_dsa.py + test_model_registry.py → 15 passed, 2 skipped,
  0 failed (stable ×3). Skips: _store conftest-stubbed (cpp parity) and the routing-parity guard.
• fp8 helpers: import GREEN; dequant_fp8_blockwise defined exactly once; ruff check+format clean.
• model_topology.cpp: clean git apply --3way (no conflict markers); clang-format 0 violations;
  a faithful port of the already-compiled b5a015a. -fsyntax-only reached the file's own body with
  no syntax errors (only a pre-existing env torch/extension.h resolution issue). Full CUDA
  pip install -e . build deferred to CI (plan-sanctioned; _store is not built in this env).
• LSP diagnostics: clean on fp8.py and test_glm_moe_dsa.py (no errors/warnings). clangd not
  installed here → C++ relied on clang-format + clean apply.
• No-duplicate checks: one SyncGlmMoeDsaMoEBlock, one glm_moe_dsa module, one each of
  dequant_fp8_blockwise/dequant_fp8_state_dict/stack_expert_scales; all DROP-list files
  (glm_moe_dsa.py, model_offload.py, constants.py, hf_config.py, C++ parallel/+
  py_archer_prefetch.cpp, test_glm_smoke.py, setup.py, README.md, test_model_registry.py)
  unmodified vs origin/dev.

Known pre-existing issue (not introduced here)

tests/python/unit/test_glm_routing.py::test_routing_parity (dev-owned, byte-identical to origin/dev)
is a flaky test: it passes in isolation / the dev-only batch / the required suite, but under one
multi-file ordering it nondeterministically produces all-NaN routing (1 failed then 39 passed on
identical re-runs). Our guard is hermetic (skips before constructing any model), so this branch
contributes no RNG/side-effects — the flake is dev-side. Out of scope to fix here; recommend hardening
that test separately.

git diff --stat origin/dev...HEAD

 core/model/model_topology.cpp                                        | 112 +++++-----
 docs/superpowers/plans/2026-08-14-glm-5.2-reconciliation.decision-log.md | 246 +++++++++++
 moe_infinity/utils/fp8.py                                            |  62 ++++
 tests/python/unit/test_glm_moe_dsa.py                                | 228 +++++++++
 4 files changed, 591 insertions(+), 57 deletions(-)