feat(sm120): vendor DeepGEMM-sm120 v0.1.0 device layer

[lucifer1004]

Summary

Replace this fork's in-tree SM120 device headers with byte-identical vendored copies of lucifer1004/DeepGEMM-sm120 v0.1.4 — a standalone single-source-of-truth repository for the SM120 device layer, shared with the nv_dev-lineage SM120 port (deepseek-ai#447). Fixes and new SM120 kernels will land there once and propagate to each fork mechanically, instead of diverging per fork.

• Nothing vendoring-specific enters this fork — no manifest, no tooling, no header markers; the vendored files are plain DeepGEMM-style headers. Provenance and drift control live on the canonical repo's side: per-tag sha256 manifest, tools/check_vendor.py --fork <path> (both directions), and a scheduled watcher that monitors registered downstream branches daily and files an issue when a tree matches no release tag.
• 13 existing headers replaced, 3 added (impls/sm120_fp8_fp4_sparse_mqa_logits.cuh, impls/sm120_padding.cuh, scheduler/sm120_gemm.cuh); the vendored files are a strict superset of this fork's versions (both lineages' fixes merged).
• Host launchers adapted to the superset kernel signatures (no behavior change for configurations this fork exercises).

Host adaptation details

• sm120_bf16_gemm / sm120_fp8_fp4_gemm_1d1d: pass the new runtime args with neutral values (EpilogueIdentity{}, stride_c_m = 0, shape_cd_m = m); template additions are trailing-defaulted.
• Masked m-grouped launchers now force swizzle_cd_mode = 0 when m % BLOCK_M != 0. This is required with the vendored kernels: they dropped the old boundary-tile scalar fallback (superseded by the nv_dev-lineage mechanism), so a boundary tile's full-tile TMA store would spill into the next group's rows. The scalar store path carries the masked row bounds (row_is_valid). Mirrors the proven nv_dev-lineage glue.
• sm120_split_k_reduce: the removed kWithAccumulation template flag maps to the runtime C operand (gmem_c = D when accumulating, else nullptr).
• Paged MQA logits: BLOCK_KV → PAGE_KV rename is semantics-preserving here (the fork already passes the real page size); the vendored kernels add within-page TMA offsets enabling PAGE_KV 128/256.

New in the vendored layer (available, inert until host glue opts in)

• CUDA ≥ 13 compile-time guard for SM120 device passes (a pre-13 ptxas was reported to silently drop block_scale).
• tensor_map_replace_global_dim_in_smem relocated into common/sm120_utils.cuh (main-lineage ptx/tma.cuh lacks it).
• Sparse MQA logits kernels, padding kernels, PAGE_KV 128/256 support, HeadSplits epilogue, K-grouped constant-stride/PSUM layouts.

Follow-up folded in (k-grouped guard)

• K-grouped GEMM (bf16 + fp8/fp4) with m % BLOCK_M != 0 could spill a boundary tile's TMA store into the next group's output slab (reproduced 5/5: next group's head rows zeroed). Both k-grouped launchers now force the scalar store epilogue for unaligned m, mirroring the nv_dev-lineage glue; regression tests test_sm120_kgroup_unaligned_m_tail_tile_isolation added to the bf16 and fp8/fp4 suites. This exposure predates the vendor migration (the old kernels had no k-grouped boundary fallback either).

Validation (sm_120a)

• tools/check_vendor.py: PASS (tree byte-identical to v0.1.4).
• SM120 suite: 48/48 passed (46 baseline + 2 k-grouped guard) in plain mode at the current head. Memcheck (0 errors) and racecheck (0 hazards) evidence is from the 50-node three-mode run at the previous head: its sanitizer executions covered exactly these tests, and the dropped nodes are test-only (the removed host checker was inert unless env-enabled). The labels-contract checker and its regression tests are deferred to upstream feat(gemm): opt-in labels-contract checker for m-grouped contiguous GEMM deepseek-ai/DeepGEMM#450 and will arrive via sync rather than as a fork-local copy.
• The vendored tree additionally passes the nv_dev-lineage SM120 suite (741 tests) on the sibling fork; the standalone repo's CI cross-compiles every header plus one representative instantiation per kernel for sm_120a against both base lineages and asserts on SASS/PTX opcodes (HMMA.16816, QMMA.*.SF.*, block_scale).

Merge-base friction

This branch merges into deepseek-ai/DeepGEMM main with zero conflicts (dry-run); all SM120-exclusive content stays in dedicated files, and VENDOR-sm120.json / tools/check_vendor.py are additive.

Follow-up folded in (v0.1.4 re-vendor)

• Re-vendored the device layer at DeepGEMM-sm120 v0.1.4 (ea3b2f8): the FP8 paged MQA logits kernel now derives BLOCK_KV = min(PAGE_KV, 64), adding PAGE_KV=32 support (DSv4.1 indexer caches mix 64- and 32-state pages; see [SM120] Support 32-state pages in FP8 paged-MQA logits #14). Device-only: this fork's host launcher still restricts paged FP8 to page 64, so page32 remains inert until host glue opts in — that is [SM120] Support 32-state pages in FP8 paged-MQA logits #14's rebased scope. Re-verified on sm_120a from a fresh JIT cache: test_sm120_mqa.py + test_sm120_fp8_fp4.py 23/23 passed.

[lucifer1004]

Force-pushed (2cf7301 → 0aecffa): dropped the fork-local labels-contract checker (DG_CHECK_CONTIGUOUS_LABELS) and its two regression nodes. Both are deferred to upstream deepseek-ai#450 and will arrive via sync, keeping this fork free of a divergent copy. Re-verified at the current head: SM120 suite 48/48 passed on sm_120a (plain mode, fresh JIT cache); the test_filter collection errors are the pre-existing helper-discovery quirk (deepseek-ai#446), not failures.