Appendix D — Bibliography

Master list, merged and deduplicated from every chapter’s ## References section (mirrored per-chapter so a reader need not hunt the back of the book). Tag conventions are defined in the README.

Muser source (pinned revision)

Quoted and cited throughout. Paths relative to the Muser repository root at pin 6d0807da (see PINNED.md).

• [crates/muser-bench/src/composite_dflash.rs:246-251] — the
• [crates/muser-bench/src/m16.rs:137-226] — the SHAPES table: dtypes,
• [crates/muser-bench/src/m16.rs:202-225] — draft shapes (fc 33280→6656,
• [crates/muser-bench/src/main.rs:304-341] — route_identity: resolved
• [crates/muser-bench/src/main.rs:305-346] — route_identity: the
• [crates/muser-bench/src/remote.rs:3-10], [:32], [:36] — the remote
• [crates/muser-bench/src/remote.rs:3-8, 32-33, 690-694] — the
• [crates/muser-bench/src/remote.rs:3-8, :33] — the 256-token +
• [crates/muser-bench/src/remote.rs:3-8, :33] — the 256-token exact
• [crates/muser-cluster/src/config.rs:10-18], :128-131 —
• [crates/muser-cluster/src/config.rs:128-131],
• [crates/muser-cluster/src/config.rs:13-18, 128-132] —
• [crates/muser-cluster/src/config.rs:41-49, 103-115, 128-131] —
• [crates/muser-cluster/src/config.rs:44-49] — enrollment-stamped
• [crates/muser-cluster/src/control.rs:1-13] — control-plane scope and
• [crates/muser-cluster/src/lib.rs:5-22] — the crate’s one-sentence
• [crates/muser-cluster/src/lib.rs:9-22] — 1× Mac + 1× GX10 launch
• [crates/muser-cluster/src/muse_sink.rs] — detached-generation contract
• [crates/muser-cluster/src/receiver.rs:108-148] — the ledger-volume
• [crates/muser-cluster/src/receiver.rs:108-150] — ledger-volume gate.
• [crates/muser-cluster/src/receiver.rs:108-206] —
• [crates/muser-cluster/src/schedule.rs:19-26, 75-90, 124-129] — NoPE
• [crates/muser-cluster/src/schedule.rs:84-157] — the span schedule:
• [crates/muser-cluster/src/security.rs:355-491] — ReplayLedger
• [crates/muser-cluster/src/security.rs] — ALPN constant (21), TLS 1.3
• [crates/muser-cluster/src/transport.rs:15-16, 19-38, 83-95] — magic,
• [crates/muser-cluster/src/verifier_v2.rs] — the unwired carried-frontier
• [crates/muser-engine/src/config.rs:169-181] — fail-closed GGUF metadata
• [crates/muser-engine/src/config.rs:286] — assert_tensor_shapes: every
• `[crates/muser-engine/src/decode.rs:1020-1026, 1331-1333, 4920-4937,
• [crates/muser-engine/src/decode.rs:1020-1030] — AcceleratorScheduler,
• [crates/muser-engine/src/decode.rs:1195-1206] — load_shared:
• [crates/muser-engine/src/decode.rs:136-139], :1209 — Metal-path
• [crates/muser-engine/src/decode.rs:213-226] —
• [crates/muser-engine/src/decode.rs:3293-3369, 3635-3666] —
• [crates/muser-engine/src/decode.rs:41-46] — llama’s nwg=32/nsg-growth
• [crates/muser-engine/src/decode.rs:53] — PREFILL_BATCH_TOKENS = 512,
• [crates/muser-engine/src/decode.rs:5432-5463] — forward_token: the
• [crates/muser-engine/src/decode.rs:6044-6091] — projection routing;
• [crates/muser-engine/src/decode.rs:6124-6279] — EncodeTarget,
• [crates/muser-engine/src/decode.rs:954-984] — MetalShared: one
• [crates/muser-engine/src/dflash.rs:1-5] — the module contract (five
• [crates/muser-engine/src/dflash/config.rs:59-61] —
• [crates/muser-engine/src/dflash/hidden.rs:44-52] — target-layer-gated
• [crates/muser-engine/src/dflash/spec.rs:16-98] — DFlashSpecStats
• [crates/muser-engine/src/dflash/weights.rs:35-66] — the dual loader
• [crates/muser-engine/src/gguf/types.rs:33-39] — NVFP4_E2M1 /
• [crates/muser-engine/src/gguf/types.rs:9-127] — GgmlType, block
• [crates/muser-engine/src/lib.rs:14] — “The pinned target artifact is
• [crates/muser-engine/src/lib.rs:14] — the pinned artifact byte size
• [crates/muser-engine/src/lib.rs:163-171] — the “no Xcode step,
• [crates/muser-engine/src/loader.rs:72-91] — weight_precision:
• [crates/muser-engine/src/loader.rs:72-91] — fail-closed lane pairing.
• [crates/muser-engine/src/metal/buffer.rs:7-14] — shared_tracked() and
• [crates/muser-engine/src/metal/context.rs:32-42] — MetalContext
• [crates/muser-engine/src/metal/context.rs:32-42] — MetalContext with
• [crates/muser-engine/src/metal/context.rs:36-39, 49-53, 111-131] — the
• [crates/muser-engine/src/metal/encode.rs:21-88] — the 66-name
• [crates/muser-engine/src/metal/encode.rs:278-280] — the pinned
• [crates/muser-engine/src/metal/encode.rs:500-523] —
• [crates/muser-engine/src/metal/encode/multicol.rs:192] — a real
• [crates/muser-engine/src/metal/encode/multicol.rs:192] — the real
• [crates/muser-engine/src/metal/encode/multicol.rs:90-132, 208-211, 458-464]
• [crates/muser-engine/src/metal/encode/norm.rs:115-116],
• [crates/muser-engine/src/metal/encode/norm.rs:236-240, 270-277] — the
• [crates/muser-engine/src/metal/encode/qkv.rs:128-227] —
• [crates/muser-engine/src/metal/encode/qkv.rs:13],
• [crates/muser-engine/src/metal/encode/qkv.rs:414-641] —
• [crates/muser-engine/src/metal/pso_cache.rs:45-49] — the
• [crates/muser-engine/src/metal/pso_cache.rs:9-50] — the in-process
• [crates/muser-engine/src/metal/residency.rs:1-107] — the
• [crates/muser-engine/src/prefill.rs:6-12] — “prefill of T tokens ≈ one
• [crates/muser-engine/src/quant/blocks.rs:100-103] — the Q5→Q8 transcode
• [crates/muser-engine/src/quant/k_block.rs:12-49] — dequant_q4_k, this
• [crates/muser-engine/src/quant/k_block.rs:169-177] — dot_q4_k_f32_llama
• [crates/muser-engine/src/quant/k_block.rs:51-105] — dequant_q5_k and
• [crates/muser-engine/src/quant/k_block/q6.rs:15-75] —
• [crates/muser-engine/src/quant/nvfp4.rs:1-6] — the pinned format
• [crates/muser-engine/src/sampling.rs:1001-1097] — the MT-pinned
• [crates/muser-engine/src/sampling.rs:1001-1097] — the source-pinned
• [crates/muser-engine/src/shaders/ferrite/batch_m16_n32.metal:59-88] —
• [crates/muser-engine/src/shaders/ferrite/rmsnorm_batch_tail.metal:1-33]
• [crates/muser-engine/src/shaders/ferrite/sigmoid_gate.metal:7-16] and
• [crates/muser-engine/src/shaders/muse_reference.metal:735-788] —
• [crates/muser-engine/src/shaders/nvfp4.metal:1-5] — the GPU format doc;
• [crates/muser-engine/src/shaders/nvfp4.metal:226+] — the
• [crates/muser-engine/src/weights.rs:163-218] — MuseWeights::open
• [crates/muser-engine/src/weights.rs:25-27] — companion-tensor suffixes;
• [crates/muser-engine/src/weights.rs:4-7] — row-contiguity / prefill
• [crates/muser-engine/tests/muse_golden.rs:97-108] — the pinned-artifact
• [crates/muser-server/src/chat_template.rs:237-261] — the release_gguf
• [crates/muser-server/src/chat_template.rs:237-261] — the Mac-side
• [crates/muser-server/src/node/artifacts.rs:188-210, 472-488] —
• [crates/muser-server/src/node/mod.rs:80-83] — MUSER_GGML_METALLIB
• [crates/muser-server/src/node/registry.rs:29-76] —
• `[crates/muser-server/src/node/smoke.rs:43-52, 451-469, 538-569,
• [crates/muser-server/src/state.rs:1666-1675] — the native+DFlash
• [crates/muser-server/src/state.rs:1666-1675],
• [crates/muser-server/src/state.rs:1667-1678] —
• [scripts/accelerator_safe.py:190-197, 202-203] — append-only journal,
• [scripts/accelerator_safe.py] — dry-run default (:35-38), lock path
• [scripts/compile_llama_metallib.sh] — revision pinning, clean-tree
• [scripts/gx10/durable_fsync_probe.py:19-22] — the standalone tail
• [scripts/gx10/llamacpp/muser_v2_send.py:54-76] — the 8 Gbps
• [scripts/gx10/llamacpp/spark_kv_export.cpp:1-30] — the integer-exact
• [scripts/gx10/README.md] — diagnostic flow, en0/en1 rule, ~9.4 Gbps
• [scripts/gx10/README.md] — the five diagnostic tools and the
• [scripts/gx10/restart_resident_producer.py] — the restart ritual:
• [scripts/gx10/tcp_probe.py], [scripts/gx10/durable_fsync_probe.py],
• [scripts/gx10/vllm/benchmark_native_prefill.py:98-102],
• [scripts/gx10/vllm/benchmark_native_prefill.py:99-102] — the native
• [scripts/gx10/vllm/benchmark_native_prefill.py:99-102],
• [scripts/gx10/vllm/muser_native_prefilld.py:2-9] — control-plane
• [scripts/gx10/vllm/muser_native_prefilld.py:514-577] — producer
• [scripts/gx10/vllm/muser_vllm/connector.py:184-285] — the CUDA side of
• [scripts/gx10/vllm/native_onboarding_identity_v1.json:57-71] — the
• [scripts/gx10/vllm/native_onboarding_identity_v1.json] — the frozen
• [scripts/gx10/vllm/resident_producer.py] — pinned vLLM commit (21),
• [scripts/gx10/vllm/supervise_resident_producer.py] — the supervisor:
• [scripts/qualify_nvfp4_fast.py:307-308] — lease enforcement below the
• [scripts/qualify_nvfp4_fast.py:333-336] — the qualifier’s matching
• [third_party/kvpack/crates/kvpack-handoff/src/canonical.rs:8-36] —
• [third_party/kvpack/crates/kvpack-handoff/src/handoff_v2.rs] —
• [third_party/kvpack/crates/kvpack-handoff/src/mac.rs] — MacKey
• scripts/compile_llama_metallib.sh — builds the pinned metallib.
• third_party/kvpack/crates/kvpack-handoff/src/handoff_v2.rs:28-91 —
• third_party/kvpack/crates/kvpack-handoff/src/manifest.rs:56-66 —
• third_party/kvpack/provenance.json — schema, upstream commit/tag/tree,
• third_party/kvpack/README.md — replay semantics (quoted), the

Muser engineering documents

Under docs/ in the Muser repository.

• [docs/decode-dispatch-gap-20260815.md] — closure accounting (common math)
• [docs/decode-dispatch-gap-20260815.md] — common-math closure accounting
• [docs/decode-dispatch-gap-20260815.md] — the 760/564 reconciliation, the
• [docs/decode-dispatch-gap-20260815.md] — the closure-count reconciliation;
• [docs/extraction-manifest.md] — silu_fast and the shader lineage from
• [docs/release-provenance.md:822-823] — logit_scale=0.196116 with
• [docs/benchmarks.md §1] — the six-depth plain matrix ratios.
• [docs/benchmarks.md §2] — natural-text wins/losses and the
• [docs/benchmarks.md §2] — verify-length conventions and the
• [docs/benchmarks.md] — §1 (35.44/35.49 parity-within-noise),
• [docs/benchmarks.md] — §Methodology (ratio convention, five-rep/CV,
• [docs/benchmarks.md] — §Methodology (repetition and floor conventions),
• [docs/campaign-review-brief-20260820.md],
• [docs/decode-dispatch-gap-20260815.md §Rejected hybrid postmortem] —
• [docs/decode-dispatch-gap-20260815.md] — closure-vs-dispatch
• [docs/decode-dispatch-gap-20260815.md] — instrument label correction;
• [docs/decode-dispatch-gap-20260815.md] — read in full for this chapter:
• [docs/decode-dispatch-gap-20260815.md] — the +196 reconciliation (104
• [docs/decode-dispatch-gap-20260815.md] — the +196-closure
• [docs/decode-dispatch-gap-20260815.md] — the 39 staging families and the
• [docs/decode-dispatch-gap-20260815.md] — the 39 staging groups and 52
• [docs/decode-dispatch-gap-20260815.md] — the 52 publication splits and
• [docs/decode-dispatch-gap-20260815.md] — the closure families of
• [docs/decode-dispatch-gap-20260815.md] — the closure reconciliation this
• [docs/decode-dispatch-gap-20260815.md] — the exactness-contract
• [docs/decode-dispatch-gap-20260815.md] — §Landed and rejected
• [docs/decode-dispatch-gap-20260815.md],
• [docs/disaggregated-prefill-sealing-plan-20260818.md §W1] — the
• [docs/disaggregated-prefill-sealing-plan-20260818.md] — §2 payload
• [docs/disaggregated-prefill-sealing-plan-20260818.md] — §2 producer
• [docs/disaggregated-prefill-sealing-plan-20260818.md] — §4 (KV-size
• [docs/disaggregated-prefill-sealing-plan-20260818.md] — §4 (W4A4
• [docs/disaggregated-prefill-sealing-plan-20260818.md] — §W0 (9.40 Gbps,
• [docs/disaggregated-prefill.md §Correctness] — refusals as passing
• [docs/disaggregated-prefill.md §What you need] — the
• [docs/disaggregated-prefill.md] — operating characteristics and the
• [docs/disaggregated-prefill.md] — roles-not-machines; the two-jobs
• [docs/documentation-truth-pass-20260815.md] — the audit table;
• [docs/extraction-manifest.md §Stage 2] — the fifteen byte-for-byte
• [docs/extraction-manifest.md] — per-file provenance for the
• [docs/extraction-manifest.md] — the ancestor’s ring-modulus hazard and
• [docs/goal-parity-ledger-2026-08.md] — the parity gates the pinned
• [docs/gx10-return-runbook-2026-08.md §2] — the live stale-generation
• [docs/gx10-return-runbook-2026-08.md] — enrolled lane constants
• [docs/kv-reuse-frontier-20260820 §1-2, §4] — the crossover and
• [docs/kvpack-merge-handoff §6] — the ~82 MB SWA / 95.7 % NoPE pacing
• [docs/kvpack-merge-handoff-20260820.md §3 D1] — the payload
• [docs/kvpack-merge-handoff-20260820.md §6] — the NoPE-trailing-edge
• [docs/kvpack-merge-handoff-20260820.md] — §3 D1 (payload 1,823,184,896
• [docs/kvpack-merge-handoff-20260820] — §1 the merge ruling, §3 F1
• [docs/kvpack-merge-handoff.md §3 D2] — the 2026-08-20 audit correcting
• [docs/kvpack.md] — the reuse ladder (Table 25.1), miss-control
• [docs/kvpack.md] — the stance (“exactness is the product”), the three
• [docs/launch-claims-review-20260824.md] — the pre-decision review
• [docs/launch-claims.md] — #2 (local matrix), #6 (the disaggregated
• [docs/launch-claims.md] — #8 (topology), #13 (producer self-recovery
• [docs/launch-claims.md] — #9 (attempt-31 exactness and wire rates).
• [docs/launch-claims.md] — #9 (wizard attempts 9 and 31, their rates
• [docs/launch-claims.md] — the register; preamble (OPERATOR REVIEW
• [docs/memory-footprint.md] — 1.827 GB/slot and 7.306 GB/four-slot KV at
• [docs/memory-footprint.md] — 96 GB M3 Ultra host, KV formula and the
• [docs/memory-footprint.md] — artifact sizes, KV formula, the
• [docs/memory-footprint.md] — batch workspace widths; chunk arithmetic.
• [docs/memory-footprint.md] — DFlash GGUF 1,631,205,312 B.
• [docs/memory-footprint.md] — KV formula, the 96 GB budget table, the
• [docs/memory-footprint.md] — the 1,024 B row formula, the one-slot /
• [docs/memory-footprint.md] — the 1,024 B/row formula and the
• [docs/memory-footprint.md] — the 1,024 B/row KV constant §16.7
• [docs/metrics-schema.md] — honesty-tag legend; nodes[] mock;
• [docs/muser-architecture.md §Context and sessions] — the five-sentence
• [docs/muser-architecture.md §Durable and remote KV] — enrollment key
• [docs/muser-architecture.md §Durable and remote KV] — the three-line
• [docs/muser-architecture.md §Durable and remote KV] — v0.1 topology,
• `[docs/muser-architecture.md §Product boundary, §Context and sessions,
• [docs/muser-architecture.md §Slots and scheduling] — the one-scheduler
• [docs/muser-architecture.md] — the lane matrix (exact-flag row:
• [docs/muser-architecture.md] — the lane matrix; “DFlash drafts are
• [docs/muser-architecture.md] — §Context and sessions (shift semantics
• [docs/muser-architecture.md] — §Slots and scheduling (shared weights,
• [docs/nvfp4-distributed-speculative-frontier-20260818.md] — the
• [docs/nvfp4-fast-lane-evidence §Measured product numbers] — the
• [docs/nvfp4-fast-lane-evidence-20260817.md §Measured product numbers] —
• [docs/nvfp4-fast-lane-evidence-20260817.md] — Fallback B disposition,
• [docs/nvfp4-fast-lane-evidence-20260817.md] — installed payload
• [docs/nvfp4-fast-lane-evidence-20260817.md] — the 6.805 tok/s
• [docs/nvfp4-fast-lane-evidence-20260817.md] — the 6.805 tok/s native
• [docs/nvfp4-fast-lane-evidence-20260817.md] — §Product route
• [docs/one-button-onboarding.md §Starting the production consumer] —
• [docs/one-button-onboarding.md] — the six stages, the security model,
• [docs/private-release.md] — the fifteen mandatory lanes (the vision
• [docs/private-release.md] — the freeze→run→readiness→seal flow
• [docs/redteam-review-campaign-brief-20260820.md §Verdict] — no
• [docs/release-provenance.md] — v0.1 scope override; §ANE v9
• crates/muser-kvpack/src/economics.rs + [docs/kvpack-economics.md] —

Campaign ledger and claims register

The measured-record instruments: docs/goal-parity-ledger-2026-08.md and docs/launch-claims.md.

• [ledger Stage B close-out] — docs/goal-parity-ledger-2026-08.md, L0
• [ledger wizard attempt 9 / nvfp4-fast-lane-evidence §Determinism] — the
• [nvfp4-fast-lane-evidence-20260817] / [ledger F-series] — the native
• [claims #10] — native NVFP4 quality gates and the published docs@65,536
• [claims #11] — docs/launch-claims.md row 11: plain Mac NVFP4 35.491
• [claims #11] — kquant 35.440 / NVFP4 35.491 tok/s scope (§6.1, §6.9).
• [claims #11] — the scope language carried in §25.3–25.4 (original
• [claims #11], [claims #12], [claims #15] — [docs/launch-claims.md]:
• [claims #12] — both delta cells, the admission rule, and the
• [claims #15], [claims #3], [claims #16], [claims #4] —
• [claims #15], [claims #6] — docs/launch-claims.md: speculative restatement scope; TTFT disaggregation scope.
• [claims #2], [claims #6], [claims #9], [claims #11], [claims #14],
• [claims #3], [claims #4], [claims #14], [claims #15],
• [claims #4], [claims #10], [claims #11] — docs/launch-claims.md:
• [claims #4], [claims #5], [claims #6], [claims #8], [claims #10],
• [claims #6], [claims #15], [claims #16] — TTFT disaggregation scope;
• [claims #6], [claims #9], [claims #13] — docs/launch-claims.md:
• [claims #9], [claims #10], [claims #11] — docs/launch-claims.md:
• [ledger "EEE A/B at 130815"] — the 570.122 s local deep-prefill mean
• [ledger "Kvpack ladder stage-5 isolated-depth verdict"] — Table 25.3’s
• [ledger "Kvpack ladder stage-6 delta-witness verdict"] +
• [ledger e426ec0 postmortem] — the retracted 65,536 `outputs_match:
• [ledger L0] — same ledger, “microbenchmark-first apparatus and the
• [ledger N2] — receiver phases constant ~0.2 s (the non-wire cost).
• [ledger P1.3] — docs/goal-parity-ledger-2026-08.md, kquant/NVFP4
• [ledger P4] — the shallow warm-hit five-sample cell, the 723.90×
• [ledger T-series "Delta-only prefill (W3)"] +
• [ledger §"EEE link ruling — operator decision (2026-08-20)"],
• [ledger §"Final GX10 campaign attempt 4"],
• [ledger §2b, 2026-08-24] — the arithmetic-ABI chase, attempts 10–31,
• [ledger §2b] — the wizard attempts 10–31 arithmetic-ABI chase (one
• [ledger §L2 Stage B verdict], [ledger §Stage B L0/L1],
• [ledger §L2 Stage B verdict], [ledger §Stage B L1],
• [ledger §N2], [ledger §N5], [ledger §P4] — the EEE collapse
• [ledger §P1.3], [ledger §2b 2026-08-24],
• [ledger §P1.3], [ledger §P1.4], [ledger §F-series remediation],
• [ledger §Stage B L0], [ledger §Stage B L1] —
• [ledger §T-series] — docs/goal-parity-ledger-2026-08.md: T0 raw
• [ledger …] — “eight-handoff deep soak” (2026-08-23); attempt-4
• [ledger …] — “Phase 4 disaggregated GX10→Mac context matrix” (the
• [ledger …] — docs/goal-parity-ledger-2026-08.md: “Stage A entry gate”
• [ledger …] — F-series remediation context and Fallback A no-go;
• [ledger …] — preamble; “Synthetic spec matrix deep-cell restatement”
• [ledger, "Phase 2 non-spec context matrix"] /
• [ledger] (docs/goal-parity-ledger-2026-08.md) — Arc 1 (0.781× → the
• [ledger] (docs/goal-parity-ledger-2026-08.md) — the campaign matrices

Retained receipts (append-only evidence volume)

Under muser-receipt://; append-only, never written by book work.

• [receipt muser-receipt://pinned-token-parity-20260814-v3/],
• [receipt final-campaign-20260823/attempt-4/soak/run-attempt-3/SOAK_VERDICT.json]
• `[receipt kvpack-ladder-20260820/attempt-10-…-stage6-delta/
• [receipt phase4-disagg-20260820/130815-g900091/] — the
• [receipt phase4-disagg-20260820/130815-g900091/] — the 1,823,184,896 B
• [receipt phase4-disagg-20260820/130815-g900091/out-p4/f-p4-text-g900091-client.json]
• [receipt wizard-validation-20260823/attempt-9-native-live-20260824T051305Z/validation-summary.json]
• [receipt …] — under muser-receipt://:
• `[receipts kvpack-ladder-20260820/attempt-9-20260822T074100Z-stage5-warmhit/

The ancestor book (pedagogical lineage)

The Ferrite inference book — structural lineage only; its numbers are ancestor context, never Muser results.

• [ferrite-book Ch 16] — the ancestor’s residual-fused o_proj matvec; the
• [ferrite-book Ch 17] — the ancestor’s SwiGLU chapter; the fused-vs-control
• [ferrite-book Ch 18] — the ancestor’s Q4_K_M mix table and the
• [ferrite-book Ch 19] — the ancestor’s LM-head chapter: “17× more rows”,
• [ferrite-book Ch 20] — the ancestor’s argmax chapter: the two-phase
• [ferrite-book Ch 5] — the ancestor Q4_K chapter; Muser’s dequant_q4_k
• [ferrite-book Ch 5] — the ancestor’s Q4_K chapter, whose
• [ferrite-book Ch 8] — the ancestor spine chapter (pedagogical lineage; its Figure 8.1 fusion list is Ferrite’s, not Muser’s).
• [ferrite-book Ch 10] — the ancestor’s RMSNorm chapter (worked example
• [ferrite-book Ch 11, Ch 12] — the ancestor’s v4/4sg GEMV chapters
• [ferrite-book Ch 13] — the ancestor’s RoPE chapter (permutation-
• [ferrite-book Ch 14] — the ancestor’s paged cache, kept as contrast
• [ferrite-book Ch 14] — the ancestor’s paged-Q8 cache (the 2×2 device
• [ferrite-book Ch 14] — the ancestor’s paged-Q8 cache: the dual-derivation
• [ferrite-book Ch 15] — the ancestor’s attention chapter (the
• [ferrite-book Ch 1] — the ancestor’s “one token, costed by hand” device
• [ferrite-book Ch 21], [ferrite-book Ch 22], `[ferrite-book
• [ferrite-book Ch 21], [ferrite-book Ch 23 §23.7] — lineage: the
• [ferrite-book Ch 23] — the roofline-flip device this chapter ports; its
• [ferrite-book Ch 24] — the ancestor’s measurement chapter: noise
• [ferrite-book Ch 25] — the ancestor’s falsification-ledger closer this
• [ferrite-book Ch 2] — the ancestor’s tape-recorder analogy and
• [ferrite-book Ch 3] — the ancestor’s unified-memory chapter: the
• [ferrite-book Ch 4] — the ancestor’s compilation chapter: the
• [ferrite-book Ch 9] — the ancestor’s CPU-lookup embedding and the

Vendor documentation

• [CUDA §streams], [CUDA §thread-hierarchy] — NVIDIA CUDA C++ Programming Guide: streams and the thread hierarchy (the CUDA-vs-Metal contrasts of Ch 29).
• [Metal-PG] — Apple, Metal Programming Guide: “Command Buffers,” “Functions,” “Pipeline States,” “Resource Objects,” and maxThreadsPerThreadgroup / threadgroup-memory limits (quoted across Part I and Ch 4).
• [Metal-SS] — Apple, Metal Shading Language Specification: “Address Spaces,” “Function Constants,” “SIMD-group Functions” (the simdgroup operations every kernel chapter relies on).
• [Metal-SS §simd-group-functions] — same specification, the SIMD-group functions section (cited where reduction order matters).

Papers

• [arxiv:1512.03385] — He et al., Deep Residual Learning (the residual stream’s origin).
• [arxiv:1706.03762] — Vaswani et al., Attention Is All You Need (the transformer this book re-derives from scratch).
• [arxiv:1710.05941] — Ramachandran et al., Searching for Activation Functions (the SwiGLU lineage).
• [arxiv:1910.07467] — Zhang & Sennrich, Root Mean Square Layer Normalization (RMSNorm; Ch 12).
• [arxiv:2002.05202] — Shazeer, GLU Variants Improve Transformer (SwiGLU; Ch 18).
• [arxiv:2104.09864] — Su et al., RoFormer (the RoPE origin paper; Ch 14).
• [arxiv:2302.01318] — Chen et al., the independent lossless speculative-decoding statement (Ch 33’s exactness contract).
• [arxiv:2305.13245] — Ainslie et al., GQA (the 16:1 bandwidth lever; Ch 9 and Ch 16).

Upstream interop targets

(no entries)