Run frontier MoE models on hardware you already own β pure C, zero deps, experts streamed from disk. Tiny engine, immense model. π¦
C
E5: validator round-1 fixes (resset mutex split, flush visibility, citations)
monotophic82d06000b5bd
diff --git a/SUMMARY.md b/SUMMARY.mdindex a4b86e36..3ccb3ae2 100644--- a/SUMMARY.md+++ b/SUMMARY.md@@ -17,12 +17,15 @@ reduction survives without the load-path tax (malloc pages never change ownershi## What changed-Everything is confined to `c/backend_metal.mm`. **No changes to `glm.c` or-`backend_metal.h`** β `coli_metal_register`/`coli_metal_unregister`'s existing signatures and-every call site in `glm.c` (expert_load, uring_load_add, qalloc, kv_alloc, map_of_fd) are-untouched; the residency-set bookkeeping lives entirely inside those two functions' existing-bodies. This is a smaller diff shape than E4, which needed new `backend_metal.h` declarations-and new `glm.c` call sites because it changed the allocation function itself.+All mechanism code is confined to `c/backend_metal.mm` β+`coli_metal_register`/`coli_metal_unregister`'s existing signatures and every call site in+`glm.c` (expert_load, uring_load_add, qalloc, kv_alloc, map_of_fd) are untouched; the+residency-set bookkeeping lives entirely inside those two functions' existing bodies. The+only `glm.c`/`backend_metal.h` touch is one instrumentation hook added in the validator+round-1 fixes (`coli_metal_resset_stats` + the gate-on-only `METAL-RESSET:` stats line in+`profile_print` β see "Validator round 1 fixes" below). Still a smaller diff shape than E4,+which needed a new alloc/free API and four new `glm.c` call-site arms because it changed the+allocation function itself.Env-gated `COLI_METAL_RESSET=1`, default OFF, runtime `@available(macOS 15.0, *)` guard witha one-line stderr fallback when requested on an older OS or when residency-set creation@@ -40,12 +43,18 @@ the `COLI_METAL_RESSET` `getenv` branch in `coli_metal_init`, so `resset_add`/`rβ one set, attached once, for the process lifetime. Failure (old OS or creation error)prints one stderr line and leaves `g_resset_enabled=false` β stock path.- **`coli_metal_register`**: after wrapping the buffer exactly as today- (`newBufferWithBytesNoCopy`), calls `resset_add(b)` under the *same* `g_slab_mtx` that- already serializes `g_slabs` mutation from parallel OMP loader threads. `resset_add` calls- `[rs addAllocation:b]` and sets a `g_resset_dirty` flag β **it does not commit**.-- **`coli_metal_unregister`**: calls `resset_remove(b)` (also under `g_slab_mtx`) *before*- clearing the `g_slabs` entry. `resset_remove` calls `[rs removeAllocation:b]` **and commits- immediately** β no batching. See UNCERTAINTIES for why this asymmetry is deliberate.+ (`newBufferWithBytesNoCopy`) and pushing the `g_slabs` entry under `g_slab_mtx` exactly as+ today, calls `resset_add(b)` **after dropping `g_slab_mtx`** but before returning.+ `resset_add` takes a dedicated `g_resset_mtx` (guarding only the set mutations and the+ dirty flag), calls `[rs addAllocation:b]` and sets `g_resset_dirty` β **it does not+ commit**. No Metal call ever runs under `g_slab_mtx` (validator round-1 fix; E4's audit+ round 2 identified mutex-over-live-Metal-call as the leading suspect for its +12s+ expert-disk regression).+- **`coli_metal_unregister`**: erases the `g_slabs` entry under `g_slab_mtx` (stashing the+ buffer), then calls `resset_remove(b)` **outside `g_slab_mtx`**, before returning.+ `resset_remove` (under `g_resset_mtx`) calls `[rs removeAllocation:b]` **and commits+ immediately** β no batching β because the caller frees the host memory right after the+ function returns. See UNCERTAINTIES for why this asymmetry is deliberate.- **`moe_submit`** (the one function whose `use` list β resolved expert weight/scale slabs βscales with LRU cache size): calls `resset_flush()` at the top (commits any pending addsfrom `resset_add`, under `g_slab_mtx`), then, if `g_resset_enabled`, **skips** the@@ -59,12 +68,13 @@ the `COLI_METAL_RESSET` `getenv` branch in `coli_metal_init`, so `resset_add`/`r### Why only `moe_submit` skips `useResource:`-Apple's own `MTLResidencySet` docs are explicit: *"Residency sets don't support hazard-tracking, so you need to account for hazards with fences and events."* (confirmed against-the actual SDK header comment and Apple's "Simplifying GPU resource management with residency-sets" guide, both read directly for this design β see UNCERTAINTIES for the exact quotes and-how they were obtained). Dropping `useResource:` therefore risks losing whatever-hazard-tracking value those calls provided. Rather than apply the residency set uniformly and+Apple's `MTLResidencySet` class reference (developer.apple.com, fetched during design on+2026-07-18) is explicit: *"Residency sets don't support hazard tracking, so you need to+account for hazards with fences and events."* The SDK header on this box+(`MTLResidencySet.h`, read directly) is **silent** on hazard tracking β the statement comes+from Apple's online documentation and adoption guide ("Simplifying GPU resource management+with residency sets"), not the header (see UNCERTAINTIES for sourcing). Dropping+`useResource:` therefore risks losing whatever hazard-tracking value those calls provided. Rather than apply the residency set uniformly andargue *in general* that hazard tracking isn't load-bearing, this diff draws the line at theone call site the mechanism history actually implicates:@@ -102,15 +112,20 @@ next `moe_submit` call, which flushes once via `resset_flush()` before it reliesfor residency.This is correct β not just fast β because of an existing invariant the codebase already-depends on for `resolve()` to work at all: a slab's `coli_metal_register` call (mutex-guarded)-always completes and releases `g_slab_mtx` before any dispatch that references that slab's+depends on for `resolve()` to work at all: a slab's `coli_metal_register` call always+completes β including its trailing `resset_add`, which runs after `g_slab_mtx` is dropped+but **before the function returns** β before any dispatch that references that slab'spointer can call `resolve()` for it (the caller in `glm.c` cannot pass a freshly-loaded-expert's pointer to a dispatch before the load β which registers it β returns). Since-`resset_flush()` also takes `g_slab_mtx`, and runs immediately before `moe_submit`'s own-`resolve()` calls in program order, any slab a given `moe_submit` invocation will resolve was-already `addAllocation:`-ed (and marked dirty) strictly before that invocation's-`resset_flush()` runs β so the flush is guaranteed to cover it, regardless of what other-threads are concurrently registering unrelated slabs.+expert's pointer to a dispatch before the load β which registers it β returns). After the+validator round-1 mutex split, the flush's synchronization runs through `g_resset_mtx`+alone: `resset_add`'s set mutation + dirty write and `resset_flush`'s dirty read + commit+are serialized by that one mutex, whose release/acquire pairs provide the memory ordering;+`g_slab_mtx` still orders the slab-table bookkeeping (register-before-resolve) exactly as on+stock. So any slab a given `moe_submit` invocation will resolve was `addAllocation:`-ed (and+marked dirty) strictly before that invocation's `resset_flush()` acquired `g_resset_mtx` β+the flush is guaranteed to cover it, regardless of what other threads are concurrently+registering unrelated slabs. The two mutexes are never held simultaneously anywhere, so no+deadlock ordering exists to maintain.`resset_remove`, by contrast, commits synchronously and immediately, with no batching,because the caller (`glm.c`, in every one of the four slab-realloc call sites, and in@@ -127,16 +142,55 @@ lifecycle wording backs this reading: "`coli_metal_register` β add allocationNo existing counter's semantics changed. `coli_metal_moe_times`/`coli_metal_moe_counts`(`g_t_setup`, `g_t_gpu`, `g_t_kernel`, `g_t_scatter`, `g_moe_ok`/`g_moe_fb`/`g_moe_experts`)are computed exactly as before β `resset_flush()` runs *before* `ts_start = mnow()` in-`moe_submit`, so its cost (whatever it is) is **outside** `g_t_setup` and therefore invisible-to the existing setup/gpu/kernel breakdown. This is a deliberate choice: it keeps the-orchestrator's A/B harness reading the same counters with the same meaning across stock/E4/E5,-but it also means the harness's existing numbers will not show E5's flush cost if it turns-out to be non-negligible β see UNCERTAINTIES. No new counters were added (no `glm.c`-`profile_print` changes), unlike E4's `METAL-HEAP: alloc fallbacks` line, because there was no-`glm.c` touch point to hang a print on without adding one β judged not worth the extra diff-surface for an experiment branch; `[METAL] residency-set: on` / the two fallback stderr lines-from `coli_metal_init` are the only new observability, sufficient to confirm which path a run-took.+`moe_submit`, so its cost is **outside** `g_t_setup`, keeping the orchestrator's A/B harness+reading the same counters with the same meaning across stock/E4/E5. The flush cost is+surfaced separately (validator round-1 fix β the original design left it invisible, a blind+spot for the battery): a dedicated `g_t_resset_flush` accumulator timed around the flush in+`moe_submit`, exported via `coli_metal_resset_stats()` (`backend_metal.h`) and printed by+`profile_print` as its own `METAL-RESSET: flush N.NNs` line β a **separate line following+the `METAL:` line, mirroring E4's `METAL-HEAP:` convention, so the existing `METAL:` line+the harness parses keeps its exact format** β printed **only when the gate is on** (the+function returns 0 when off), so stock output stays byte-identical. The register-side+`resset_add`/`resset_remove` costs have no dedicated counter: they run inside the engine's+existing expert-load wait accounting (the `t_ewait` window in `glm.c`), noted in a comment+at `resset_add`, so a load-path regression from set bookkeeping would already show in the+existing disk/wait numbers. `[METAL] residency-set: on` / the two fallback stderr lines from+`coli_metal_init` confirm which path a run took.++## Validator round 1 fixes++1. **REQUIRED, Metal calls hoisted out of `g_slab_mtx`** (`backend_metal.mm`): the original+ design ran `addAllocation:`/`removeAllocation:`/`commit` while holding `g_slab_mtx`, the+ lock the parallel OMP loader threads contend on β structurally identical to the+ mutex-over-live-Metal-call shape E4's audit round 2 identified as the leading suspect for+ its replicated +12s expert-disk regression, and the SDK header notes commit on a resident+ set tries to make resources resident "instantly" (real synchronous work; this set is+ resident from startup since it is queue-attached for the process lifetime). Fixed by+ introducing a dedicated `g_resset_mtx` guarding only the set mutations + dirty flag;+ `g_slabs` push/erase stays under `g_slab_mtx` exactly as stock; the two mutexes are never+ held together. The registerβflushβresolve happens-before argument is preserved β see the+ updated "Deferred-commit design" section and the comment at `resset_add`.+2. **REQUIRED, false citations corrected** (this file + the `moe_submit` commit message,+ rewritten pre-push): the original text attributed the hazard-tracking and thread-safety+ statements to the SDK header (`MTLResidencySet.h`), which is in fact silent on both+ topics. The statements come from Apple's **online** `MTLResidencySet` class reference and+ the "Simplifying GPU resource management with residency sets" adoption guide (both+ fetched 2026-07-18 during design). All attributions now name the actual source; where a+ claim rests on design reasoning rather than documentation, it is labeled as such.+3. **REQUIRED, flush cost made harness-visible**: `g_t_resset_flush` ++ `coli_metal_resset_stats()` + the gate-on-only `METAL-RESSET:` line in `profile_print` β+ see "Instrumentation parity" above.+4. **NOTE β pre-existing fslab OOM-unwind bug becomes a longer-lived artifact under E5**+ (no code change on this branch, per the fix round's scoping): `expert_load`'s fslab OOM+ path (`c/glm.c` ~1868β1873 on this base) frees `s->slab` via `compat_aligned_free`+ **without** `coli_metal_unregister` β on stock this leaves a stale `g_slabs` entry whose+ GPU exposure ends with the last command buffer that declared it; under E5 the buffer is+ additionally a **permanent residency-set member** referencing freed host memory until+ some later realloc of the same slot unregisters by pointer, a strictly longer-lived+ exposure than stock's transient per-CB one. When E5 graduates to the upstream PR, it must+ carry the one-line unregister-before-free fix; E4's branch has the reference+ implementation (`6753225`, "validator round-1 fixes (heap-create latch, unwind+ unregister)").## Per-seam differences vs E4@@ -163,7 +217,9 @@ took.fix-plan's "Determinism side-finding").5. **`[METAL] residency-set: on` line present in stderr** at flag-on startup, and absent(or the OS<15/create-failed fallback line) otherwise β cheap sanity check that a run- actually exercised the intended path before trusting its numbers.+ actually exercised the intended path before trusting its numbers. Also read the+ **`METAL-RESSET: flush` line** (gate-on only): if that number is large, the deferred+ set-commit cost is eating the stall win from the dispatch side.6. If the hypothesis holds (E5 stall β E4, E5 load-path β stock, identical output), E5 becomesthe upstream PR candidate and must include the cap-default recalibration flagged in PR#386's CURRENT-STATE CALIBRATION markers, per the spec's validation plan.@@ -173,7 +229,9 @@ took.`cd c && make glm METAL=1` and a separate explicit `-Wall -Wextra` compile of`backend_metal.mm` (the Makefile's `METALXX` line does not itself pass `-Wall -Wextra`, so"clean under `-Wall -Wextra`" was checked with those flags added explicitly), plus-`cd c && make glm` (plain, non-Metal β unaffected, since this diff never touches `glm.c`), and+`cd c && make glm` (plain, non-Metal β the one `glm.c` touch, the `METAL-RESSET` stats line,+is inside the pre-existing `#ifdef COLI_METAL` arm of `profile_print`, so the plain build+compiles none of it), and`make metal-test` (existing synthetic kernel-correctness unit test β no model, no`glm52_i4/`, random weights β run once with `COLI_METAL_RESSET` unset and once with`COLI_METAL_RESSET=1` to numerically exercise `coli_metal_register`/`moe_submit`'s changed@@ -188,14 +246,16 @@ with the existing queue/command-buffer structure, or something unverifiable withmodel run β flagged per the task's hard requirement.**1. **The central design risk: skipping `useResource:` in `moe_submit` gives up Metal's- automatic hazard tracking for that buffer set.** Apple's `MTLResidencySet.h` header- (read directly from this box's SDK at- `/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk/.../Headers/MTLResidencySet.h`)- documents the protocol only in terms of residency; Apple's own "Simplifying GPU resource- management with residency sets" guide states plainly: *"You don't need to call- `useResource`/`useHeap`... for allocations in a residency set,"* and separately, the- `MTLResidencySet` class reference states *"Residency sets don't support hazard tracking,- so you need to account for hazards with fences and events."* I reasoned through every+ automatic hazard tracking for that buffer set.** Sourcing (corrected in validator round+ 1): the SDK header on this box+ (`/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk/.../Headers/MTLResidencySet.h`,+ read directly) documents the protocol only in terms of residency and says nothing about+ hazard tracking either way; the two operative statements are from Apple's **online**+ documentation (fetched 2026-07-18): the "Simplifying GPU resource management with+ residency sets" adoption guide β *"You don't need to call `useResource`/`useHeap`... for+ allocations in a residency set"* β and the `MTLResidencySet` class reference β+ *"Residency sets don't support hazard tracking, so you need to account for hazards with+ fences and events."* I reasoned through everycode path that touches `moe_submit`'s `use` buffers (read-only, indirectly referenced,never concurrently written, freed only after the engine's own slot lifecycle guaranteesno outstanding async reference) and concluded removing `useResource:` there specifically@@ -205,30 +265,33 @@ model run β flagged per the task's hard requirement.**been) that existed before. **This is the #1 thing to watch for md5 divergence on**, andthe reason the scope was deliberately narrowed to `moe_submit` alone rather than applieduniformly.-2. **`resset_add`/`resset_remove`/`resset_flush` all run under `g_slab_mtx`, serializing- concurrent OMP loader threads' residency-set mutations against each other and against- dispatch's flush.** `MTLResidencySet`'s own header says *"all methods are non-threadsafe"*- (confirmed directly from the SDK header), so this serialization is required for- correctness, not optional β but it is structurally the same shape as the bug E4's- audit-round-2 found and fixed (mutex held across a live Metal call, serializing loader- threads during warmup fan-out, suspected root cause of E4's +12s regression). Apple's- guide frames `addAllocation:`/`commit` as lightweight bookkeeping relative to the actual- (deferred, async) page-in work ("Metal makes allocations resident when you call `commit()`- on the first command buffer using the set" β implying `commit()` itself doesn't- synchronously page anything in), which is why I judged holding the mutex across these- calls acceptable unlike E4's `newBufferWithLength:` (a real allocation call). **This is- unverified without profiling a loaded run** β if `commit()` or `addAllocation:` turns out- to be synchronously expensive on this hardware/OS build, this could reproduce E4's- expert-disk-load regression through a different code path, defeating E5's entire premise- (decoupling residency from the load path). Orchestrator: this is the single most important- number to check E5's load-path timing against stock, not just against E4.+2. **Residency-set mutations are serialized under a dedicated `g_resset_mtx` (validator+ round-1 fix β originally they ran under `g_slab_mtx`, the E4-regression shape; no Metal+ call runs under the slab lock anymore).** The serialization itself is kept as required+ for correctness: Apple's online `MTLResidencySet` class reference states the set's+ *"methods aren't thread-safe"* (the SDK header contains no thread-safety statement either+ way β citation corrected in round 1; the online doc is the source). What remains+ **unverified without profiling a loaded run** is the *cost* of the calls themselves:+ `resset_remove`'s synchronous `commit` runs inside `coli_metal_unregister` on the+ loader path (its cost lands in the existing `t_ewait` accounting), and the SDK header+ says commit on a resident set tries to make added/removed resources resident/non-resident+ *"instantly"* β real synchronous work, since this set is resident from startup+ (queue-attached for the process lifetime). If `commit()`/`addAllocation:` turn out+ expensive on this hardware/OS build, the load path degrades through set bookkeeping+ rather than mutex contention β a different, now-decoupled failure mode, but the same+ symptom as E4's regression. Orchestrator: check E5's load-path timing against stock, not+ just against E4, and read the new `METAL-RESSET: flush` line for the dispatch-side share.3. **`resset_flush()`'s cost sits outside `g_t_setup`/the `moe_times` breakdown** (it runs- before `ts_start = mnow()`), by design, to keep the harness's existing counters meaningful- β but this also means if `commit()` is expensive, it will show up as *general* wall-clock- slowdown / tok/s regression without a corresponding line in the existing instrumentation- pointing at it. No new counter was added for it (see "Instrumentation parity" above) to- avoid a `glm.c` touch; if E5 numbers look off without an obvious cause in the existing- breakdown, `resset_flush`/`commit()` cost is the first place to add a throwaway probe.+ before `ts_start = mnow()`), by design, to keep the harness's existing counters+ meaningful β and, since validator round 1, it is **no longer invisible**: the+ `g_t_resset_flush` accumulator surfaces it as the gate-on-only `METAL-RESSET: flush`+ line (see "Instrumentation parity"). Residual blind spots: (a) the accumulator is a+ plain double written from `moe_submit` on the engine thread, matching the existing+ `g_t_setup` convention β if `moe_submit` were ever called from multiple threads+ concurrently, both counters would be equally wrong; (b) the register-side+ `resset_add`/`resset_remove` costs have no dedicated counter and are only visible+ blended into the existing `t_ewait`/disk-wait numbers (comment at `resset_add` says so)+ β a fine-grained attribution would need a throwaway probe.4. **`initialCapacity = 4096` on the `MTLResidencySetDescriptor` is an unverified guess.**It's documented as a presize hint only (no correctness effect either way), chosen to be"clearly larger than the permanent-weight-tensor + KV-cache + plausible cap16 LRU-slab@@ -273,3 +336,10 @@ model run β flagged per the task's hard requirement.**high-confidence. What is **not** independently verified is runtime behavior beyond whatthe docs state and what the synthetic unit test exercises β no substitute for theorchestrator's real cap-sweep battery.+10. **Pre-existing fslab OOM-unwind bug has a strictly longer-lived exposure window under+ E5** β see "Validator round 1 fixes" item 4. Not fixed on this branch (out of the+ experiment's scope per the fix round), but the upstream PR built from E5 **must** carry+ the one-line unregister-before-free fix (reference implementation on E4's branch,+ commit `6753225`). The bug is only reachable through the fslab-OOM path (allocation+ failure mid-load), so it cannot affect the orchestrator's controlled A/B runs at sane+ RAM headroom β flagged for the PR, not for the battery.diff --git a/c/backend_metal.h b/c/backend_metal.hindex 4383b3b6..e6e24dd8 100644--- a/c/backend_metal.h+++ b/c/backend_metal.h@@ -98,6 +98,10 @@ int coli_metal_attn_decode(const float *x,void coli_metal_moe_counts(uint64_t *ok, uint64_t *fb, uint64_t *experts);void coli_metal_moe_times(double *setup, double *gpu, double *scatter);double coli_metal_moe_kernel_time(void);+/* E5 (COLI_METAL_RESSET=1): returns 1 when the queue-attached residency set is active and+ * writes the cumulative seconds moe_submit spent committing pending set adds -- a cost that+ * sits OUTSIDE the setup/gpu/scatter breakdown above. Returns 0 (and writes 0) when off. */+int coli_metal_resset_stats(double *flush_s);/** Batched routed-expert SwiGLU for one MoE block, in ONE command buffer.diff --git a/c/backend_metal.mm b/c/backend_metal.mmindex 9c10533d..be7dae8d 100644--- a/c/backend_metal.mm+++ b/c/backend_metal.mm@@ -270,23 +270,36 @@ static std::mutex g_slab_mtx; // expert_load registers slabs from parallel Ope// @available(macOS 15.0, *) guards below, keeping -Wunguarded-availability clean.static id g_resset_obj;static bool g_resset_enabled; // COLI_METAL_RESSET=1, macOS 15+, and creation succeeded-static bool g_resset_dirty; // addAllocation: calls pending commit; g_slab_mtx-guarded--// Add a just-wrapped buffer to the set. Deferred commit (batches an OMP loader burst into-// one commit at the next moe_submit instead of one per slab): correct because every-// register/unregister caller holds g_slab_mtx here, and resolve() (which every dispatch-// calls before touching a slab) takes the same mutex -- a slab moe_submit's resolve() can-// see was necessarily added, and marked dirty, before moe_submit's own resset_flush() ran.-static void resset_add(id<MTLBuffer> b) { // caller holds g_slab_mtx+static bool g_resset_dirty; // addAllocation: calls pending commit; g_resset_mtx-guarded+// Set mutations + dirty flag get their OWN mutex, never held together with g_slab_mtx: no+// live Metal call may run under the slab lock the parallel OMP loader threads contend on+// (E4's audit round 2 found exactly that shape -- mutex over a live Metal call -- as the+// leading suspect for its +12s expert-disk regression). g_slab_mtx keeps guarding g_slabs+// bookkeeping only, exactly as on stock.+static std::mutex g_resset_mtx;+static double g_t_resset_flush; // sec committing pending adds in moe_submit (gate on only)++// Add a just-wrapped buffer to the set; commit deferred (an OMP loader burst batches into+// one commit at the next moe_submit instead of one per slab). Called by coli_metal_register+// after it drops g_slab_mtx but before it returns -- and the engine cannot dispatch an+// expert before the load that registers its slab returns, so any slab a given moe_submit+// can resolve() was added (and marked dirty) under g_resset_mtx strictly before that+// moe_submit's resset_flush() acquired the same mutex: the flush covers it. The slab-table+// ordering itself (register-before-resolve) is unchanged and stays under g_slab_mtx.+// Cost lands in the caller's existing expert-load accounting (t_ewait window in glm.c);+// no separate counter for the add/remove side.+static void resset_add(id<MTLBuffer> b) {if (!g_resset_enabled) return;+ std::lock_guard<std::mutex> lk(g_resset_mtx);if (@available(macOS 15.0, *)) { [(id<MTLResidencySet>)g_resset_obj addAllocation:b]; g_resset_dirty = true; }}// Remove + commit immediately, NOT deferred: the caller frees the underlying host memory// right after coli_metal_unregister returns, so the removal must be applied before that --// an uncommitted-but-still-resident allocation pointing at freed memory is a use-after-free-// risk the GPU could act on.-static void resset_remove(id<MTLBuffer> b) { // caller holds g_slab_mtx+// risk the GPU could act on. Also runs outside g_slab_mtx (see g_resset_mtx above).+static void resset_remove(id<MTLBuffer> b) {if (!g_resset_enabled) return;+ std::lock_guard<std::mutex> lk(g_resset_mtx);if (@available(macOS 15.0, *)) {id<MTLResidencySet> rs = (id<MTLResidencySet>)g_resset_obj;[rs removeAllocation:b]; [rs commit];@@ -294,14 +307,23 @@ static void resset_remove(id<MTLBuffer> b) { // caller holds g_slab_mtxg_resset_dirty = false; // commit above also flushes any pending adds}// Flush pending adds before moe_submit relies on the set alone for residency -- the only-// caller that skips per-buffer useResource: (see moe_submit below).+// caller that skips per-buffer useResource: (see moe_submit below). Takes g_resset_mtx+// only, never g_slab_mtx; the happens-before argument lives at resset_add above.static void resset_flush() {if (!g_resset_enabled) return;- std::lock_guard<std::mutex> lk(g_slab_mtx);+ std::lock_guard<std::mutex> lk(g_resset_mtx);if (!g_resset_dirty) return;if (@available(macOS 15.0, *)) { [(id<MTLResidencySet>)g_resset_obj commit]; }g_resset_dirty = false;}+// Harness visibility for the flush cost, which sits OUTSIDE the moe_times setup/gpu+// breakdown (timed around resset_flush in moe_submit, before ts_start). Returns whether+// the set is active so glm.c prints the METAL-RESSET line only when the gate is on --+// stock output stays byte-identical.+extern "C" int coli_metal_resset_stats(double *flush_s) {+ if (flush_s) *flush_s = g_t_resset_flush;+ return g_resset_enabled ? 1 : 0;+}// Persistent scratch buffers (grow-only) for the MoE pipeline.static id<MTLBuffer> g_gg, g_uu, g_hh, g_xg; static size_t g_gg_cap, g_uu_cap, g_hh_cap, g_xg_cap;@@ -377,17 +399,22 @@ extern "C" void coli_metal_register(void *base, size_t len) {id<MTLBuffer> b = [g_dev newBufferWithBytesNoCopy:base length:lenoptions:g_res_opts deallocator:nil];if (!b) return;- std::lock_guard<std::mutex> lk(g_slab_mtx); // called from parallel expert_load threads- for (auto &s : g_slabs) if (s.base == base) { s.len = len; s.buf = b; resset_add(b); return; }- g_slabs.push_back({base, len, b});- resset_add(b);+ {+ std::lock_guard<std::mutex> lk(g_slab_mtx); // called from parallel expert_load threads+ bool found = false;+ for (auto &s : g_slabs) if (s.base == base) { s.len = len; s.buf = b; found = true; break; }+ if (!found) g_slabs.push_back({base, len, b});+ }+ resset_add(b); // E5: outside g_slab_mtx (no Metal call under the slab lock); still before return}extern "C" void coli_metal_unregister(void *base) {- std::lock_guard<std::mutex> lk(g_slab_mtx);- for (size_t i=0;i<g_slabs.size();i++) if (g_slabs[i].base==base) {- resset_remove(g_slabs[i].buf); // must land before the caller frees base- g_slabs[i].buf=nil; g_slabs.erase(g_slabs.begin()+i); return;+ id<MTLBuffer> b = nil;+ {+ std::lock_guard<std::mutex> lk(g_slab_mtx);+ for (size_t i=0;i<g_slabs.size();i++) if (g_slabs[i].base==base) {+ b = g_slabs[i].buf; g_slabs[i].buf=nil; g_slabs.erase(g_slabs.begin()+i); break; }}+ if (b) resset_remove(b); // E5: outside g_slab_mtx; commits before the caller frees base}// Resolve a host pointer inside a registered slab to (buffer, gpuAddress). Returns nil if unknown.static id<MTLBuffer> resolve(const void *p, uint64_t *addr) {@@ -743,7 +770,9 @@ static id<MTLCommandBuffer> moe_submit(int nb, int D, int Iinter, int fmt,const float *xg, const int *xoff, const int *nr, int R,id<MTLBuffer> xg_buf, id<MTLBuffer> gg_buf, id<MTLBuffer> uu_buf, id<MTLBuffer> hh_buf) {if (!g_dev || (fmt != 1 && fmt != 2)) return nil;- resset_flush(); // E5: commit any pending slab adds before we may skip useResource: below+ if (g_resset_enabled) { // E5: commit any pending slab adds before we may skip useResource:+ double t0 = mnow(); resset_flush(); g_t_resset_flush += mnow() - t0; // METAL-RESSET line+ }double ts_start = mnow();std::vector<uint64_t> ag(nb),au(nb),ad(nb),sgv(nb),suv(nb),sdv(nb);std::vector<id<MTLBuffer>> use; use.reserve(nb*2);diff --git a/c/colibri.c b/c/colibri.cindex 03eb878b..d08615ab 100644--- a/c/colibri.c+++ b/c/colibri.c@@ -3801,7 +3801,9 @@ static void profile_print(Model *m, double elapsed){if(aok){ double ks=0,gs=0; coli_metal_attn_lat(&ks,&gs);printf("METAL-ATTN: layer GPU %llu | gpu-wall %.2fs (kernel %.2fs | cpu-sched %.2fs gpu-sched %.2fs)\n",(unsigned long long)aok,aw,ak,ks,gs); } }printf("METAL: blocchi GPU %llu | fallback CPU %llu | expert su GPU %llu | setup %.2fs gpu-wall %.2fs (kernel %.2fs) scatter %.2fs\n",- (unsigned long long)ok,(unsigned long long)fb,(unsigned long long)ex,su,gp,coli_metal_moe_kernel_time(),sc); }+ (unsigned long long)ok,(unsigned long long)fb,(unsigned long long)ex,su,gp,coli_metal_moe_kernel_time(),sc);+ { double rsf=0; if(coli_metal_resset_stats(&rsf)) /* E5: printed only when the gate is on */+ printf("METAL-RESSET: flush %.2fs (residency-set commit in moe_submit, outside setup/gpu-wall)\n",rsf); } }#endif}