feat: device_fit governor calc + resident-override wire + K3 fork bump - #2109
Merged
Conversation
…nto canary Advances the vendored llama.cpp fork 30 commits (clean FF over canary's stale 66594cc3f): container-serve resident-override (LLAMA_RESIDENT_OVERRIDE), the rung-2 ResidencyCache plan-file consumer, the score-hint/generation-bias actuator, PagerCaptureEvent emit, fit-device --reserve-gb. Makes canary USE the K3 misfit-serving stack (measured 0.33 tok/s WASTE-parity on a 32GB card). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…M5's replay
Live GGML_MOE_TRACE_FILE slice (12B records: u64 tkey + u32 e) + the reverse
table so BanditPlanController recovers (layer,expert): tkey=FNV-1a of
blk.{layer}.ffn_{gate,up,down}_exps.weight, e=within-layer expert idx, expert
identity=(layer,e) deduped across the 3 matrices. RUN-1 static-pin datum input.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
The actual std-only Rust prototypes written against live K3 traces this session: trace_replay (recency beats LFU 3-4x), predictor (offline learned-decay +5pts held-out), online_predictor (bandit 49.8 vs 47.8 best-fixed on non-stationary), self_optimize (joint speed×quality). These are the faithful-port source for the learned policy behind TierPolicy (continuum-core expert_tier_policy.rs, #276). Numbers are properties of these exact constants + reward math — reproduce before improving. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
The major GPU speedup: promote hot experts to persistent VRAM so decode's hot path is GPU-native (zero fetch, zero copy). 3 increments (copy-skip -> VRAM hot cache -> pipeline), the 32GB rate-distortion constraint (imatrix-enabled resident shrink frees VRAM for the hot set), measured per-increment via k3-bench. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…vs 2 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…DeviceUploadFetcher Mechanism is the existing (buft,fetcher)-generic ResidencyCache; a VRAM cache = same class + device buft + host->device fetcher. 3 small parameterized pieces (DeviceUploadFetcher, instantiate w/ GGML_MOE_VRAM_CACHE_GB, seam hook). Stats only tune params -> zero mechanism rework. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…rnor seam) Diagnoses the hardcoded-cache overcommit that collapsed K3 fetch bandwidth (40GB pinned + mmap = 95.9GB on 63GB -> pagefile thrash -> 205 MB/s -> 0.027 tok/s) and lays out the clean architecture: governor owns the residency budget net of the model's mmap footprint, plan-file is the one wire, ResidencyCache is pure mechanism. Governor-interface sections marked [M5 OWNS] for her to edit. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…f-mmap is explicit arithmetic; plan_file.budget_bytes is the lease wire) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo
…path Records the BigMama measurements feeding M5's #287 derivation (non-cache ~56GB, per-token working set 5.5GB, governed budget ~6GB, fetch recovers to 2.5GB/s at fit), the now-complete C++ cache mechanism (enable-from-plan, grow, shrink), and the three-piece graduated path to serving/load kimi-k3 (catalog row + serving-lane MoE launch + #287) replacing the rigged .bat. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
start-server.sh now provisions the full Windows CUDA build env before the cargo builds (the cargo/nvcc path had none, unlike the vcvars-wrapped llama cmake): import MSVC via vswhere->VS2022-14.4x + a .bat env dump (cl.exe for nvcc), pin CMAKE to the manifest install, force CMAKE_GENERATOR=Ninja (the VS18-2026 auto- pick is undefined in cmake 3.30), add the Windows SDK bin (mt.exe/rc.exe), select a complete CUDA toolkit + CUDA_PATH (a provisioning split left cuda-env with 0 import libs vs cuda-13.2's 12), and RUSTFLAGS -L for pocket-tts (which emits no link-search) while re-carrying +crt-static so the /MT GPU stack still links. Portability: expert_container.rs + commands/capacity.rs used Unix-only std::os::unix::fs::FileExt::read_exact_at. Add crate::platform_io::pread_exact (unix read_exact_at / windows seek_read loop) - one place for positioned reads. Build validated (npm start exit 0, continuum-core lib clean). A separate runtime hot-loop on the #2088 core at startup is tracked apart from this build fix. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
Pure calc the governor uses to fit a streaming-MoE's RESIDENT (non-expert) tier to a device VRAM budget and reconcile it with the expert tier on ONE budget — fixing the double-count where the expert pager was handed the full VRAM ceiling while resident silently ate most of it. Partition (in order): compute reserve -> resident (Native | device-fit Override | Unfittable) -> sufficient-context KV -> everything left = hot-expert VRAM budget (maximized: more on-GPU experts, fewer streams). Context is derived + clamped, never hand-picked. Artifact resolver injected (no hardcoded paths). Standalone-validated 7/7; M5 wires it into the daemon spawn path + launch (ServingTarget.resident_override) per the K3 sprint split. Refs #29 #31 #36. Arch-confirmed on real K3 UD-IQ2 (93 blk/896 exp/top-16). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
Wire foundation for the governor's device_fit plan: ServingTarget carries resident_override: Option<PathBuf>, and the launcher exports it as LLAMA_RESIDENT_OVERRIDE so llama.cpp sources the precision-shrunk RESIDENT (non-expert) tensors from the device-fit GGUF (all offloaded to GPU) while the primary streams experts. All builders updated; defaults None (resident serves as-shipped, no behavior change) until compute_resident_override + the resolve-or-generate resolver (#35) land next. In-crate validated. Refs #29 #36. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
M5's per-layer KV accessors (n_head_kv_il + n_embd_head_{k,v}_il, continuum #238)
+ graph reconciliation. The K3 engine now builds against these — enables the
device_fit resident-override serve + honest per-layer K3 KV sizing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
Contributor
Author
|
M5 review (formal approve blocked — same GitHub identity): APPROVED. Merge when your compute step is ready, or merge this foundation now — it's inert ( What's right, notably: Two notes for the
🤖 Generated with Claude Code |
# Conflicts: # tools/scripts/start-server.sh
The governor now DECIDES the resident source per serve: compute_resident_override derives resident_bytes (weights - expert_bytes_total) vs the governed VRAM ceiling via capacity::device_fit, and sets ServingTarget.resident_override. A dense/small model fits native (None); a >VRAM-resident MoE (K3) resolves a cached device-fit override that fits, else Unfittable → route to grid / generate (#35), glass-boxed. resolve_device_fit_override (model_registry::artifacts): looks up a per-user device-fit cache convention (<storage_root>/device-fit/<id>/) + a resident-bytes sidecar; returns the override only when its resident fits the usable budget. No hardcoded paths; generation/HF discovery is #35. The resident-fit decision turns only on resident_bytes vs budget — per-layer KV (#2107 ModelCapabilities) drives the context/expert split elsewhere, so KV is not consulted here. Refs #29 #35 #36. In-crate validated. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
Contributor
Author
|
Review of ec1b541 (compute_resident_override) — approve, with two binding follow-ups before the K3 serve consumes downstream outputs: ✅ The resident-fit verdict is correct and honest: Native/Override/Unfittable turns only on resident_bytes vs (budget − reserve), trust-but-verify rejects oversized overrides, Unfittable is loud.
Neither blocks the LiveUploadPager run — go. 🎯 |
…naged like VRAM/RAM Joel: 'like vram and memory, this contention has to be managed between cold storage and nvme.' Design: NVMe is a governed HOT-SERVING tier (a ResourcePool, same TrackedDir + evict_at_least machinery as CargoTargetPool), whose eviction = MIGRATE frozen/duplicate artifacts to the Cold drive, not a manual rm. Serving asks ensure_hot_resident(model); composes with device_fit's Unfittable one tier down (VRAM). Corrects the DriveRole bug: Cold (HDD) is FROZEN storage, never the per-token streaming tier (HDD = unservable). Dissolves today's K3 container disk fight: the C: IQ2 is a verified duplicate of the D: copy -> governor migrates it off NVMe -> container fits, no human deletes anything. Refs #12 #36. Design for M5's system_resources lane. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
… for the storage tier The gate the NVMe serving-tier eviction (#302) consults before dropping a frozen GGUF: is an IDENTICAL twin already on cold storage? is_structural_twin (pure) = same shard count + per-shard name + size, zero-byte shards never match. scan_shards + find_cold_twin are the thin fs layer. Never drop an NVMe artifact without a VERIFIED cold twin (dropping 662GB on a path guess is the failure this guards). Standalone-validated 5/5. Composes with device_fit + M5's NvmeServingTierPool. Refs #12 #36. Design: STORAGE-SERVING-TIER-GOVERNOR.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…rve wired Both halves of the DirContainerFetcher wire (BigMama fetcher + moe_pick_fetcher branch 175ac9d6a; M5 caller-side encode + record_bytes reader c6469d5). Serving now reads the aligned per-layer container (GGML_MOE_CONTAINER) instead of the scattered raw GGUF — the honest ~2.6GB/s path. Retires the built-not-wired ContainerFetcher. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
# Conflicts: # core/continuum-core/src/modules/serving_daemon.rs
…coverage measurement Merge fix: vision_sidecar's ServingTarget was missing resident_override (added by #29). Plus a measurement test that drains the real K3 routed-access fixture through the #282 predictive instrument and prints repeat_recall / predicted_delta / schedulable_coverage — the go/no-go for the LiveUploadPager predictive pipeline (H2D/token = (1 - coverage) x ~11GB). Prints, never asserts (real routing sample). NOTE: can't run on windows-msvc (pre-existing cargo-test Unix-socket block, ipc/mod.rs); runs on M5's Mac. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…, --once, --budget-slots) The moe-pager-driver gains an offline replay mode so any completed GGML_MOE_TRACE_FILE can be scored on any box (windows-msvc clean by crate constraint), not just tailed live next to a serve: - --synth-layers N: synthesize the tkey->layer map from layer count alone (TkeyTable::for_layers, the same zero-config seam MoeTraceTail owns) instead of requiring an operator tkey-to-layer-matrix.json. - --once: exit when the trace stops growing (EOF) and print a SUMMARY line with mean DECODE-token serving hit = warm schedulable coverage. - --budget-slots N: override the predictor residency budget (default auto = first token x1.5) to measure the coverage-vs-free-VRAM curve (the device-fit tradeoff). Measured on BigMama run2.trace (302 warm decode tokens): bandit coverage 13.8% @250 slots -> 51.3% @2000 -> 65.7% @4024, beating naive last-N recency by +7-9pts at matched VRAM. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…ver for K3 VDD offline measurement (cooccur-ceiling bin) on the real warm serve trace (run2.trace, 122 held-out decode tokens, 11102 layer-steps): cross_layer_cooccur_hit 0.159 (adjacent-layer noisy-OR) recency_same_layer_hit 0.403 (last token, same layer) structure beyond recency -0.244 cooccur_recall_on_recency_misses 0.112 (11824/106058) Adjacent-layer co-occurrence predicts <half what plain recency does, and recovers only 11% of the experts recency misses (~base rate). K3 expert routing has no exploitable cross-layer structure — the CrossLayerExpert- Predictor prefetch lever is not worth wiring (saves the ggml pass-id capture slice). Recency-family residency (the bandit EMA curve) is THE signal; the only lever that lifts K3 is freeing VRAM (device-fit shrink) so residency coverage can reach the measured 51%. Caveat: adjacent-layer, one workload trace. Wider-predecessor noisy-OR would regress toward the frequency baseline (which underperforms recency), so a large lift is unlikely — but not measured here. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…ache half (#23) Pins the fork at the DeviceUploadFetcher wiring (my half of the LiveUpload- Pager H2D-kill). Off unless GGML_MOE_VRAM_CACHE_GB / plan device_budget_bytes enables it; host serving path byte-for-byte unchanged. M5's expert-loop D2D half lands next on the same seam. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
… + tier manifest (#40) Enables the device-fit division: produce a small resident override per precision tier + a (tier_label, resident_bytes) sidecar the governor reads to co-optimize the VRAM split. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
#3) The second control rung above the pager's DecayBandit. The pager decides WHICH experts stay resident (reward=hit-rate, cheap, online). This decides HOW TO DIVIDE the card — resident (non-expert) weights vs expert cache — to MAXIMIZE tok/s. That reward (actual tok/s) is EXPENSIVE (a serve), so naive online RL flails; the fix is SIM-WARM-START: predict tok/s per division OFFLINE from the measured coverage curve, then a slow bandit refines each arm from real measured tok/s. - CoverageModel: piecewise-linear coverage(slots) over MEASURED points (k3_measured() = the trace-replay curve); saturates, never extrapolates up. - predict_tok_s: coverage -> (1-coverage)*experts/token*expert_bytes H2D -> t_token -> tok/s. Higher coverage -> less H2D -> faster (the load-bearing property, tested). - feasible_divisions: tier catalog (from --resident-only manifests) x HardwareBudget -> cache budget/slots per tier; drops VRAM-overflow tiers. - DivisionBandit: warm_start from the predictor; observe(tier, measured_tok_s) overrides the prior on first serve then EMAs — the expensive reward spent only on the arm actually run. Policy lives here (windows-clean, 4 tests pass); serving_daemon actuates it (M5's #2: discover manifests, feed catalog+budget+live tok/s, apply the chosen {resident_tier, device_budget_bytes} to the plan). Fractal control law: pager (experts<->hit-rate) -> this (VRAM split<->tok/s) -> grid. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…e (clamp to free VRAM + null-buffer guard, #23) Testing convicted the segfault as VRAM oversubscription (K3 33GB resident + env cache on a 32GB card, cudaMalloc lazy-VMM deferred fault). Fix: clamp device budget to measured free VRAM (mine) + M5's D2D null-buffer guard. Device cache now disables safely where there's no room (K3) and works where there is (V4-Flash); can't crash from any budget source. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…lash curve Feeds the real BigMama RTX 5090 --n-cpu-moe sweep (DeepSeek-V4-Flash UD-IQ2_M) into DivisionBandit: 0 resident=1.39, 8 resident=1.69, 14 resident=1.68 tok/s. Asserts the bandit converges on the SATURATION KNEE (8 layers), not max residency — 8->14 layers buys nothing at +11GB VRAM. Encodes the measured finding that the governor must learn 'minimal static residency + max device cache', the freed VRAM belonging to the recency cache (#43), not to over-pinned static layers. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
Measured V4-Flash device-cache coverage curve (5090, GGML_MOE_VRAM_CACHE_GB sweep): 6GB/992slots=1.80, 12GB/1985=3.10, 22GB/3630=2.96 tok/s, all 100% hit. tok/s is NON-MONOTONIC in budget: undersized churns, 12GB is the plateau knee, 22GB is no better (100% hit but O(slots) reserve_slot eviction scan). predict_tok_s's monotonic prior would pick 22GB; only the measured reward lands on 12GB — which frees ~20GB of a 32GB card for co-resident lanes. Pins the invariant that the governor must not oversize the cache and starve other models. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
…sync restore + enum fix Includes the prefetch host_visible guard (THE #43 crash fix, validated 3.05 tok/s V4-Flash device cache on the 5090), M5's async cpy_tensor_async restore, and the moe-pack quant-enum fix. A fresh parent build now includes the un-crashable device cache instead of the pre-fix pin. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc
# Conflicts: # core/continuum-core/src/capacity/expert_container.rs # core/continuum-core/src/commands/capacity.rs
Contributor
Author
|
Merged via admin override. The one red check was a KNOWN-FLAKY timing test — |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The current windows-cuda-serving lane (supersedes the stale PR #2056, which predates the canary K3-expert modules).
Composes with M5's #2107 (per-layer KV) + #2108 (uniform_offload_required GDN guard). Arch-confirmed on real K3.
🤖 Generated with Claude Code