lfm2: draft with LiquidAI's DSpark drafter (--drafter) by BlivionIaG · Pull Request #1391 · ashhart/TensorFold

BlivionIaG · 2026-10-26T11:37:32Z

Current source a87815724d8fc37597d9b9a65e78dea7ffec7fb6 includes pinned main 7651ce64f, including Ash’s merge of #1381. Not ready for merge: full performance qualification is pending on this source. Conflicts are resolved: GitHub reports this head MERGEABLE. Exact-head CI:9/9 SUCCESS. All three workflow runs independently report this exact head SHA; earlier green runs are not relabelled.

Scope and review changes

Adds the opt-in LiquidAI DSpark drafter through --drafter, with released BF16 weights by default and explicit 8-bit quantization. The existing depth rule remains the headless scheduler. The rejected shared/lone guards and fixed scheduler modes are absent. Existing refusal/error-path fixes remain: unsupported rank/activation/rope and 4-bit admission, failed wider-lane refusal, held-pick validation, and scratch/warm cleanup.

All stages use the existing allocator with the drafter’s measured stream-count/depth grid. Only neural blocks pay head cost; copy and forced rows still occupy verify capacity. Initial holds wait for a funded grant. Split holds pay for every physical pass, and settled GPU work is attributed at the existing wait without adding synchronization. Normal copy pricing is restored in drafter arms without changing the no-drafter configuration.

Prompt and verified taps remain owned before the target overwrites them, and are projected on a hold, snapshot or stage overflow. Zero-grant rounds retain contiguous context for later drafts; absorption is not simply skipped. No rejected stream-count guard is reintroduced.

Current source-bound checks

CPU lane consumer suite: 318 passed; 1 skipped; 0 failed. The skip is Windows-only.
HTTP-serving executable plus affected family/metadata scope: Build Summary: 72/72 steps succeeded; 130/130 tests passed, using zig build native test-lfm2 test-lfm2-dspark test-native test-draft-metal-host test-drafter-skip-metal-host -Dcpu=apple_m1 -Dversion=1.2.0 -j1.
Six real-model native scenarios on this PR’s own source: BF16 and 8-bit at 16-, 48- and 2,144-token contexts. They exercise owned context, overflow, snapshot/resume, release during a shared pending hold and physical pass splitting. Existing fp32-reference near-tie criteria are unchanged; these are not final-target token or throughput qualification.
Formatting passes. Exact pinned-main checker: 424 inherited findings, zero new; checker/gates unchanged.
Own-head HTTP replay passed: 24 exact cells / 30 oracle records, code/chat, greedy/sampled, 1/4 streams, 64 tokens, with three clean exits. Solo, draft:false, concurrent and cross-arm target token hashes match. This is correctness evidence, not the 512-token performance gate.
The clean top-head QUIET1 run completed but failed32/32throughputcomparisons. The remaining superseded R2 sequence was deliberately stopped with one wrapperSIGTERM; it drained/reaped its servers. Partial1391receipts remain unqualified;1392/1393didnotstart. Isolated R3 capture/pricing fixes are being implemented and runtime-checked first. All four final updated heads still require the complete unchanged512-token/code/chat/greedy/sampled/1/4/8/16stream gate. No throughput pass or selected-cell waiver.

Full performance gate and historical provenance

The prior own-head complete matrix belongs to 7225d5478ad1eed4df8c683312d271ad6c7a4a69: 144/144 cells and 306/306 oracle records exact, three balanced rotations and nine clean shutdowns, but 32/32 speed comparisons failed. It is not a measurement of the new head.

The latest maintainer M5 measurement supersedes the older “1–2% tradeoff” interpretation: every 4-/8-stream cell was slower in BF16 and 8-bit, with additional sampled 1-/16-stream losses. The required new gate remains same-head no-drafter versus BF16 and 8-bit, code/chat, greedy/sampled, 1/4/8/16 streams, 512-token replies, three balanced rotations, exact tokens and clean exits. Every one of the 32 per-arm medians must reach its own plain baseline. No tolerance, pooled cells, selected greedy-only rows or short smoke substitutes.

Historical correctness, long-context and cancellation receipts and restack/device/load-memory receipts retain their original source/device provenance. They are not relabelled as measurements of this head. New short/native checks do not replace them. Fresh source-bound startup peaks now compare the new confidence/allocation heads in BF16 and 8-bit, two balanced orders, before any generation; they are not serving-memory or minimum-RAM claims.

Part of #1374. No force-push, assistant PR merge or CI-gate weakening. Ash merged #1381 himself; the remaining chain still needs performance qualification and maintainer review.

ashhart · 2026-10-26T11:44:43Z

A small request, @BlivionIaG: this description links more than 20 issues and PRs, which is over Agent Traffic Control's limit, so ATC drops the PR from its board and collision checks (and marks every check incomplete for everyone). Could you trim the related-work list to the issues this PR actually fixes or depends on (20 at most)? Thanks.

ashhart

Thanks @BlivionIaG, a lot of careful work in this stack. We mapped all 31 commits to their PRs and reviewed each on current main. For the DSpark drafter (this PR and its descendants):

Loading changes the released BF16 weights. dspark_weights.zig quantizes each dense matrix to 4/8-bit codes at startup (8-bit by default), and #1393 still discards the original. Target verification protects final tokens, but the engine doesn't create lossy copies of a checkpoint at serving time. Please run the released BF16 matrices, or refuse the format. Also honour each advertised --drafter-bits mode or refuse it by name (the native adapter accepts the setting and then uses 8 regardless).
4-bit mode reads past its buffer. With 4 bits the loader packs markov_w2 at that width, but lfm2_markov_scan (lfm2_dspark.metal) reads a byte per coefficient and the launcher passes no bit width. At 128000x256 the code buffer is 16,384,000 bytes and the shader can address byte 32,767,999 (found statically, not run on a GPU). Please refuse that mode before load or implement it correctly.
The held (asynchronous) path drops draft failure flags: the synchronous path refuses a NaN pick or invalid ID (DraftNonfinite), but readPicks after an unwaited hold ignores Pick.nonfinite and maps an invalid ID to 0, as does the GPU-held copy. Native serving enables that path, so please carry the flag through the fence and add planted NaN/Inf/invalid-ID controls for both paths.
Contract proof: the new operators report 0 planted controls and no recorded compiler/launch limits (the limit declarations use current-device checks with offline = false). Please add the gates, red controls, served shapes and records.
Config admission doesn't check hidden_act before the fixed SiLU path, and an explicit scheduled verify without a confidence head silently becomes depth mode; both should refuse by name.
The FP32 reference hard-codes a 6-bit target embedding and its host matmul joins workers only after all spawns succeed.
The device receipts are mostly on older bases; we'll need alternating A/B on current main per landing head (the stack's own notes report a 1.8-2.7% four-stream drop on Qwen3.8-27B + DFlash2 from the shared-lane changes, which needs explaining, not dismissing).

BlivionIaG · 2026-10-26T20:03:45Z

Thanks @ashhart. Pushed 35e9255c0 (main merged in).

Weights: with no flag, or --drafter-bits 0, the drafter runs the released bf16 matrices as they are. --drafter-bits 8 is an explicit option that quantizes them to 8-bit codes at load, named in --help and the startup log. 4 is refused on LFM2, 8 on other families. On the M5 Ultra, bf16 is 2.6% to 4.8% slower than 8-bit on one stream with about the same acceptance; numbers are in the body.
4-bit mode: refused by name before load.
Held path: a NaN, Inf or invalid pick now fails with DraftNonfinite on both paths, with planted controls that fail without the fix.
Contract: the drafter's kernels (bf16 ones included) have planted controls and recorded limits. M4 Pro --only lfm2: 11 kernels pass, 17 of 17 controls red as planted. M5 Ultra at the stack head 2e5966f68: --only lfm2 11 kernels pass gates 1 to 4, 17 of 17 controls red as planted, 0 limit problems.
hidden_act other than silu is refused by name. There is no verify flag any more, so no scheduled-without-head case.
The fp32 reference reads the target embedding bits from the checkpoint and joins its workers on every path.
A/B on current main: in lanes: a lone stream holds drafts only while its drafted rounds measure faster #1382. Qwen3.8-27B with DFlash2 at 4 streams, M5, 6 rotated rounds, median of round medians: main 322.0 tok/s, a second build of the same main commit 318.8 (-1.0%), lane head 318.9 (-1.0%). Identical source differs as much as the lane head does, and the spread inside one cell is 304 to 333, so I did not reproduce the 1.8% to 2.7% gap; it follows the build or the session, not the code.

Chips: an M3 Pro ran the stack head (@jschmied's receipt on lfm2: a cheaper DSpark hold (one concurrent encoder) #1394): tf-lfm2-dspark 9 of 9 drafts equal the fp32 reference, drafted replies equal "draft": false, 0 host allocations. M1 and M2: not run.

ashhart · 2026-10-27T14:00:01Z

Thanks @BlivionIaG. A lot of this is fixed, and fixed well:

by default the drafter runs the released bf16 matrices as stored (only the q and k rows and their norm gains are
reordered for the half-split RoPE);
4-bit is refused before load, so the markov_w2 overread cannot happen;
the held path carries failures: lfm2_draft_ids sets a flag that runHeld reads before any cache commits, and
readPicks refuses a NaN or out-of-range pick as the waited path does;
a hidden_act other than silu is refused, and the fp32 reference reads the embedding width and joins its workers
on every path;
every new kernel has a launch record and planted controls (11 kernels, 17 controls at the stack head).

What still needs to change:

Drafting off by stream count. The drafter's facts set shared_guard = true, so served DSpark runs lanes: a shared round drafts only while drafting measures faster (Model.shared_guard) #1379's
guard, which we are not taking (see lanes: a shared round drafts only while drafting measures faster (Model.shared_guard) #1379). Please remove it and let the lane core's pricing decide (lanes: an external drafter's chain chances price verify rows; holds follow the grants and pay for their pass #1381 with the
confidence head, the depth rule without it). With the guard on, the stack head (bf16) is +17% to +23% on one
stream, about even at 4 and 8 streams, and 1.2% slower at 16. Without the guard, --drafter must still be at
least as fast as no drafter at every admitted width and context; we would also like drafting to pay at 4 to 16
streams, so the hold's cost there is the thing to work on.
The timed fix is missing on this head. series/dspark-drafter carries lanes: a shared round drafts only while drafting measures faster (Model.shared_guard) #1379's first three commits but not
4fa02928a, so with Config.timed = false the guard still keeps adaptive state; our table-only check fails here
(expected 0 guard entries, found 1) and passes on lfm2: the DSpark confidence head schedules the verify #1392 to lfm2: a cheaper DSpark hold (one concurrent encoder) #1394. Removing the guard settles this too.
Drafter.init frees owned before walking it on an error. Errdefers run in reverse order, so
d.owned.deinit(gpa) runs first and the for (d.owned.items) loop then reads freed memory. A failure in scratch()
or warm() therefore walks freed memory (a minimal test of the same pattern segfaults under Debug) instead of
refusing --drafter cleanly. One errdefer fixes it:
errdefer { for (d.owned.items) |b| b.deinit(); d.owned.deinit(gpa); }. Please add a test that fails a scratch
allocation or warm().
Config admission. markov_rank: 512 is admitted, but lfm2_markov_scan stages W1's row in
threadgroup float e[256], so any rank over 256 writes past it; rope_theta: 1e100 becomes an infinite f32 and
passes theta > 0. Please add rank > 256 to the UnsupportedDraftShape check and !std.math.isFinite(theta) to
BadDraftConfig, with both cases in the refusal test.
Quantizer test. 64 * bits is computed in u8 at dspark_quant.zig:88 and overflows at 8 bits. The test passes
only because zig build test builds the engine tests at -Doptimize=fast; under Debug it panics.
64 * @as(usize, bits) / 8 fixes it.
--drafter after a failed wide check. If the local wider-lanes check fails, Metal.init keeps one-row rounds
and Drafted.facts then gives zero drafts, so the server loads the drafter and silently never drafts. Please refuse
--drafter by name when the backend's width is 1.
--drafter-bits 8 (LFM2.5-24B-A2B on Metal: drafts with LiquidAI's DSpark drafter (proposal and gate) #1374, Question 2). An opt-in 8-bit drafter is fine. It only changes what the drafter
proposes; the target's verify keeps every token exact. That matches --drafter-bits for the other families' drafters
on main. Please make sure drafted still equals plain with it, and keep bf16 the default.

Once these are in and the stack is rebased on current main without #1379 and #1382, we will run drafted against plain,
concurrent against solo and the speed cells on our Mac.

… holds follow the grants A drafter that gives each held chain draft's chance (Drafter.VTable.chances) may set Facts.chances. Drafted.facts then sets Model.draft_probabilities and Model.granted_depth, so the window allocator prices each stream's verify rows by those chances, as it does a head's node chances. With Facts.confidence_floor it instead verifies the drafts at least that likely (a fixed window when 0). With granted_depth a stream's next hold is as deep as its last grant plus a slack: one more draft when the grant took every held draft, none when it stopped short of them (Stream.grant_short), and after a grant of none a probe of draft_slack drafts at the depth rule's probe interval. Facts.slack sets one slack for all three. A shared round's granted depths are cut so the drafting streams' blocks fit one hold pass of batch_rows. When the shared-round guard starts a drafted stretch, its first round grants every stream one draft, so the stretch measures drafted rounds rather than plain ones left by grants of none. All of it is behind Facts.chances and Model.granted_depth, off by default; windows.allocate records grant_short only for granted depths. Tests: chances price the chain's rows and drafted replies equal one-token decode alone and at 1, 4 and 8 shared streams, with the scheduled verify ragged across streams and verifying fewer rows than a fixed window; the facts reach the lane config; granted depths, their slacks and the one-pass cut; holds as wide as the grants.

…py, and holds kept only while they pay The depth rule asks one head draft for a stream with a copy ahead, and a shared round's granted holds were priced only by their verify rows. A drafter whose hold runs as a pass of its own (Facts.copy_first) pays that pass for the one draft, and the copy round then drops it; Model.copy_first now asks it for none. A drafter may also give its hold pass's timed cost by blocks and depth (Facts.hold_costs, priced by hold_price.zig). With granted depths a shared round then keeps granted holds, most expected drafts a row first, only while each raises the next round's expected tokens a ms over its verify rows and the pass; a probe after a grant of none keeps its hold. A stream's last grant records the drafts it expected to keep (Stream.grant_gain). Both are off by default (copy_first false, hold_costs empty). Tests: with copy_first no hold is asked for a copy round and replies equal one-token decode at 1 and 3 streams; a hold's cost between and past its timed passes; granted holds kept while they pay, none for a dear pass, and a probe's hold kept.

Bound the reused ordering buffer by batch_rows, and cover reuse with an allocator that refuses further allocations. Interpolate between the measured depths surrounding the requested depth, including nearest-end extrapolation. Query a stateful copy proposer once for a granted-depth budget decision.

The lane core can put an external drafter in front of a target. The drafter reads the target's hidden rows after some layers (its taps). LFM2-MoE had no taps, so it could draft only with the core's own heads. prepare_features names the layers. From then on, each forward copies each row's output of those layers into the scratch (lfm2_tap, a plain copy of bf16 values). features() gives a stream's rows of the latest verify or one-token round. prefill_features hands each prompt chunk's rows to the drafter before the chunk's end is kept for reuse. The prompt's last row is kept beside the stream's caches for its first draft. A drafting stream resumes only a kept state that also holds its drafter's context (Snap.draft). Without a drafter nothing changes: no taps are kept and every forward runs as before. lfm2_tap registers with the kernel contract (metal_lfm2_tap.zig): the third of five taps' columns at 1, 10, 32 and 2,048 rows must hold the layer's bits and the other columns what the call found. Gate 5 checks its launch against every Apple chip we serve and this GPU. test-lfm2-dspark checks the same copy through the family's own kernels.

…ontracts

BlivionIaG · 2026-10-29T22:49:12Z

@ashhart Rebuilt on current main d33dcfc76, including clean #1390 taps and #1381 pricing, at 8e9b548fd. The shared guard is gone. The cleanup now frees buffers before their list, with scratch/warm allocator-failure coverage; config refuses rank >256 and nonfinite theta; quantizer arithmetic is widened under Debug; width-one qualification refuses --drafter explicitly. BF16 remains the default, 8-bit is opt-in, and 4-bit is refused. Held-path failure flags and kernel contracts are retained. Local scoped checks: 353 pass / 3 skip, zero new lean findings. The downstream 1–2% slowdown at four/eight streams remains disclosed rather than claimed fixed or waived; the prior larger 16-stream gain does not satisfy your all-width performance requirement. No M5 was accessed for this restack. Fresh CI is running.

…r growth

# Conflicts: # CHANGELOG.md

BlivionIaG · 2026-11-01T09:35:02Z

Complete own-head qualification for 7225d5478ad1eed4df8c683312d271ad6c7a4a69: FAIL, exit 3. 32/32 BF16/8-bit throughput comparisons are below plain; no negative cell is waived. This PR is not performance-qualified or merge-ready.

The entire required matrix completed: 144/144 cells, 306/306 oracle records, 512-token replies, three balanced rotations, code/chat, greedy/sampled, 1/4/8/16 streams. Plain/BF16/8-bit use the exact same frozen binary for this PR, not the cumulative stack's binary. Exactness and scenario failures are zero; all nine servers exited gracefully with code 0 and no forced kill. Hardware: M4 Pro / 48 GiB, not an M5 measurement.

All comparisons below use per-cell medians of the three rotations:

Prompt	T	Streams	Plain tok/s	BF16 tok/s (Δ)	8-bit tok/s (Δ)
code	0	1	169.225	166.329 (-1.711%)	167.602 (-0.959%)
code	0	4	290.884	285.060 (-2.002%)	287.635 (-1.117%)
code	0	8	316.876	311.954 (-1.553%)	313.038 (-1.211%)
code	0	16	348.471	346.653 (-0.522%)	347.344 (-0.323%)
code	1	1	114.154	111.987 (-1.899%)	112.140 (-1.765%)
code	1	4	201.299	194.038 (-3.607%)	179.183 (-10.987%)
code	1	8	238.444	229.804 (-3.624%)	219.413 (-7.981%)
code	1	16	289.057	280.148 (-3.082%)	268.752 (-7.025%)
chat	0	1	119.275	116.534 (-2.297%)	118.346 (-0.778%)
chat	0	4	246.822	243.431 (-1.374%)	240.966 (-2.373%)
chat	0	8	284.255	278.193 (-2.133%)	267.208 (-5.997%)
chat	0	16	343.463	340.026 (-1.000%)	323.986 (-5.671%)
chat	1	1	117.267	114.788 (-2.114%)	111.779 (-4.680%)
chat	1	4	212.095	199.494 (-5.941%)	199.056 (-6.148%)
chat	1	8	249.325	223.360 (-10.414%)	233.821 (-6.219%)
chat	1	16	305.601	300.796 (-1.572%)	289.550 (-5.252%)

Raw results remain immutable and source-bound; this failed run will not be pooled with other heads or a new candidate. The remaining landing heads continue sequentially without competing local builds or GPU/model work. Any optimization must preserve the accepted depth-rule/headless behavior, complete drafter context and fresh-zero reservations; no stream-count disabling or weakened gate.

All nine exact-head CI checks and local native/scoped checks passed, but that does not close the performance gate. Historical receipts retain their original provenance.

Source/binary/model/driver provenance, raw-file digests, every shutdown

{
  "source_revision": "7225d5478ad1eed4df8c683312d271ad6c7a4a69",
  "binary_sha256": "bbbf1882660dc4a7742ebb6ebf5bbe46ebbafcf81a6625d5b24463efa6390987",
  "pinned_main": "b17fb507af5ec1da599c5381b6990d9067e6a965",
  "target_revision": "667408d494bde191c2cb9dd4f4549827816b9275",
  "drafter_revision": "e052e5b1dd103c38cdd4f8ddc068d7dea4586f55",
  "driver_source_revision": "a1b4f2b94d223d37b1ad0a3a4a67490859307aa8",
  "driver_sha256": {
    "bench_lfm2_dspark.py": "928a1c1b6751e184243c48d187e552b622874da0ec6168a101e3d636664e3e97",
    "bench_lfm2_dspark_verdict.py": "c30d4c902979acaab150440bcf76d95492394e2abe0fe720903cb5f3090881ce",
    "bench_concurrent.py": "3234c3edfc71725d9a66f1a6485f1bc43870a3ee0c797f90ac1e7ef9a6ae74e2",
    "bench_evidence.py": "54c28474f0804dbd120e3fb5859be0d1667ef21ca502880d0bed098ec9bb42ff",
    "bench_auth.py": "1896120b70d0df05646870a252697e20ce84c63e09dcba88b2a4e5a55ccb3c5b",
    "bench_metrics.py": "26f4c3ecbb03df4f12e24d18840fea989d9ef8c63117ca11eb944ef670458a27",
    "bench_timing.py": "158f636ef0b3bed460ce19ef4872b6d8607ff7a8590562d06bae8bd55e6f44ed"
  },
  "environment": {
    "os": "macOS-26.5.1-arm64-arm-64bit-Mach-O",
    "os_build": "25F80",
    "machine": "arm64",
    "hw_model": "Mac16,7",
    "cpu": "Apple M4 Pro",
    "memory_bytes": "51539607552",
    "python": "3.14.7"
  },
  "raw_file_sha256": {
    "metadata.json": "3de29c232bc6b2640bf11b3ac69ba754d0b42bfc6afde071bb220edb873a7b78",
    "requests.jsonl": "bb9ec110d5893d98f074013e9b03f525adef0384ab075d5db42e0e396652ac5b",
    "cells.jsonl": "0f591859dfa676eba796ff5f4b0f31cdf2ccf247ef8d015e3570c3f2b88f73c2",
    "events.jsonl": "53f7ad4180e91521a572c34113ecf0575eec1b297dc5f7366b1d1b865de6d203",
    "summary.json": "b49998fc29cfb1a3d8d1fc22dcd31c40a646a065d14f98f8a20e960cba5d3704"
  },
  "shutdowns": [
    {
      "event": "shutdown",
      "at": "2026-11-01T08:17:06Z",
      "label": "r0-plain",
      "forced_kill": false,
      "seconds": 0.6,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T08:26:11Z",
      "label": "r0-bf16",
      "forced_kill": false,
      "seconds": 0.4,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T08:35:25Z",
      "label": "r0-bits8",
      "forced_kill": false,
      "seconds": 0.4,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T08:46:01Z",
      "label": "r1-bf16",
      "forced_kill": false,
      "seconds": 0.4,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T08:56:05Z",
      "label": "r1-bits8",
      "forced_kill": false,
      "seconds": 0.3,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T09:05:18Z",
      "label": "r1-plain",
      "forced_kill": false,
      "seconds": 0.4,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T09:14:18Z",
      "label": "r2-bits8",
      "forced_kill": false,
      "seconds": 0.4,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T09:23:15Z",
      "label": "r2-plain",
      "forced_kill": false,
      "seconds": 0.5,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T09:32:20Z",
      "label": "r2-bf16",
      "forced_kill": false,
      "seconds": 0.3,
      "returncode": 0
    }
  ]
}

Isolated source candidate: full pass costs and settled GPU timing, capacity-clipped staging, and behavioral regressions. Runtime checks and performance qualification are pending; the frozen matrix job is unchanged.

# Conflicts: # docs/receipts/pr1381-review/README.md # zig/src/core/lanes/config.zig # zig/src/core/lanes/depth.zig # zig/src/core/lanes/drafted.zig # zig/src/core/lanes/drafter.zig # zig/src/core/lanes/granted_reservation_test.zig # zig/src/core/lanes/hold_price.zig # zig/src/core/lanes/shared_drafts.zig

Source-only integration checkpoint. Compilation, native execution, and complete own-head performance qualification remain pending.

BlivionIaG · 2026-11-01T14:46:06Z

Own-head correctness replay completed on the newly published source below. Code/chat, greedy/sampled, one/four streams,64tokens; plain, BF16 and8-bit. All target hashes agree with solo, draft:false, concurrent and cross-arm oracles; three normal exits, no forced kill. The six native scenarios independently cover short/larger/overflow context, snapshot/resume and release during a pending shared hold.

Not performance qualification. The separate complete512-token,1/4/8/16-stream, three-balanced-rotation gate is now running on immutable own-head HTTP server binaries. It must pass all32 individual drafter-versus-plain medians; earlier source failures and historical receipts are not relabelled. CI is still independently pending/observed, not waived.

{
  "scope": "Own-head code/chat, greedy/sampled, 1/4 streams, 64 tokens; not performance qualification",
  "source_revision": "a87815724d8fc37597d9b9a65e78dea7ffec7fb6",
  "server_binary_sha256": "9da45aff1ccc9d3599ec1c6e84aae3d6d8af613b751128a283f5c6fb25afe62a",
  "pinned_main": "7651ce64fcfdcde9325721218534c50eb71c15ee",
  "cells": 24,
  "oracles": 30,
  "exactness_failures": 0,
  "clean_shutdowns": [
    {
      "event": "shutdown",
      "at": "2026-11-01T14:26:21Z",
      "label": "economics-r2-1391-plain",
      "forced_kill": false,
      "seconds": 0.5,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T14:26:53Z",
      "label": "economics-r2-1391-bf16",
      "forced_kill": false,
      "seconds": 0.5,
      "returncode": 0
    },
    {
      "event": "shutdown",
      "at": "2026-11-01T14:27:25Z",
      "label": "economics-r2-1391-bits8",
      "forced_kill": false,
      "seconds": 0.7,
      "returncode": 0
    }
  ],
  "core": "318 passed; 1 skipped; 0 failed.",
  "serving_family_scope": "Build Summary: 72/72 steps succeeded; 130/130 tests passed",
  "native_device_scenarios": 6,
  "native_context_lengths": [
    16,
    48,
    2144
  ],
  "drafter_bits": [
    0,
    8
  ],
  "new_lean_findings": 0,
  "inherited_lean_findings": 424
}

This was referenced Oct 26, 2026

lfm2: the DSpark confidence head schedules the verify #1392

Open

LFM2.5-24B-A2B on Metal: drafts with LiquidAI's DSpark drafter (proposal and gate) #1374

Open

This was referenced Oct 26, 2026

lanes: Drafted passes held windows to a target that reads them itself #1378

Closed

lanes: a shared round drafts only while drafting measures faster (Model.shared_guard) #1379

Closed

lfm2: keep a drafter's tapped rows on every forward #1390

Merged

ashhart requested changes Oct 26, 2026

View reviewed changes

This was referenced Oct 26, 2026

lfm2: the DSpark drafter loads one matrix at a time and keeps a lane in one buffer #1393

Open

lfm2: a cheaper DSpark hold (one concurrent encoder) #1394

Open

BlivionIaG marked this pull request as ready for review October 26, 2026 21:37

akol1 mentioned this pull request Oct 27, 2026

core: add the row-split pair seam and split simd_qmm over K #1432

Merged

7 of 9 tasks

BobClawblaw mentioned this pull request Oct 27, 2026

moss_image: MOSS-Image-Base's transformer on the Metal runtime (new family, denoiser) #1457

Closed

9 tasks done

jayleaton mentioned this pull request Oct 29, 2026

A pull request skill for contributor agents #1595

Draft

4 tasks done

BlivionIaG added 8 commits October 29, 2026 23:18

docs: record rebased lane pricing and DSpark hold retention

81e073a

lfm2 taps: retain launch limits, negative controls and current main c…

4156e9e

…ontracts

Merge commit '81e073a2b' into fix/restack-1391

fc6726e

lfm2: restore DSpark on main with BF16 defaults and review safety fixes

8e9b548

BlivionIaG force-pushed the series/dspark-drafter branch from 35e9255 to 8e9b548 Compare October 29, 2026 22:45

BlivionIaG added 2 commits October 31, 2026 12:03

Merge current main into lane pricing

2fea951

Merge current main into DSpark drafter

e560dab

davidtai mentioned this pull request Oct 31, 2026

hip: Qwen3.8-Flash-Next on gfx1151, a reference branch for review (to be split) #1667

Draft

BlivionIaG added 2 commits November 1, 2026 01:13

Merge current main while retaining LFM2 drafter precision options

dc63e5a

Merge current main into lane pricing review

06bc1b0

BlivionIaG added 8 commits November 1, 2026 01:19

Make DSpark allocation-failure controls deterministic across allocato…

8824808

…r growth

lanes: compose grant-priced holds with current shared reservations

ac9b868

docs: clarify shared hold probe exceptions

8fcaca5

Merge fix/pr1381-review into restack-1391

1b26d45

# Conflicts: # CHANGELOG.md

merge: integrate current main b17fb50 for standalone pricing

28da502

Merge standalone pricing into restack 1391

913b437

lanes: subtract priced padded holds from observed fixed overhead

b833c60

merge: carry priced-hold overhead attribution into PR1391

7225d54

BlivionIaG added 7 commits November 1, 2026 12:26

lfm2: price physical drafter passes and retain owned context

0d554b5

Isolated source candidate: full pass costs and settled GPU timing, capacity-clipped staging, and behavioral regressions. Runtime checks and performance qualification are pending; the frozen matrix job is unchanged.

lanes: avoid shadowing the per-stream head mask

feb8b3c

Source-only integration checkpoint. Compilation, native execution, and complete own-head performance qualification remain pending.

Fix hold allocation test array syntax

395a409

lanes: document physical pricing with concise invariant comments

49125d8

lfm2: document funded physical passes and retained context

2cfa805

lfm2: exercise owned context overflow resume and pending release

a878157

Navigation Menu

Repository navigation

lfm2: draft with LiquidAI's DSpark drafter (--drafter) - #1391

lfm2: draft with LiquidAI's DSpark drafter (--drafter)#1391
BlivionIaG wants to merge 27 commits into
ashhart:mainfrom
BlivionIaG:series/dspark-drafter

BlivionIaG commented Oct 26, 2026 •

edited

Loading

Uh oh!

ashhart commented Oct 26, 2026

Uh oh!

ashhart left a comment

Uh oh!

BlivionIaG commented Oct 26, 2026 •

edited

Loading

Uh oh!

ashhart commented Oct 27, 2026

Uh oh!

BlivionIaG commented Oct 29, 2026

Uh oh!

BlivionIaG commented Nov 1, 2026

Uh oh!

BlivionIaG commented Nov 1, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

BlivionIaG commented Oct 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Scope and review changes

Current source-bound checks

Full performance gate and historical provenance

Uh oh!

ashhart commented Oct 26, 2026

Uh oh!

ashhart left a comment

Choose a reason for hiding this comment

Uh oh!

BlivionIaG commented Oct 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ashhart commented Oct 27, 2026

Uh oh!

BlivionIaG commented Oct 29, 2026

Uh oh!

BlivionIaG commented Nov 1, 2026

Uh oh!

BlivionIaG commented Nov 1, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

BlivionIaG commented Oct 26, 2026 •

edited

Loading

BlivionIaG commented Oct 26, 2026 •

edited

Loading