LocalAI/backend/cpp/llama-cpp-localai-paged/docs/PARITY_HANDOFF.md

PARITY_HANDOFF: how to pick up the GB10 vLLM-parity work

2026-06-30 update: this handoff is now historical procedure, not the active verdict. The GB10 investigation was reopened in GB10_PARITY_REOPEN_SPEC.md and GB10_PARITY_PHASE0_RESULTS.md, with Phase 6 serving-nsys evidence and the active follow-up plans under docs/superpowers/plans/. Use those files for the current state before relying on the older "closed" conclusion below.

Audience: an agent with zero prior context who has been told to "continue the GB10 vLLM-parity investigation" on the llama-cpp-localai-paged backend.

This file is the operational how-to. It is the companion to VLLM_PARITY_FINAL.md, which is the why / authoritative record ("never re-litigate"). If the two ever disagree on a fact, VLLM_PARITY_FINAL.md and the bench artifacts it cites win; this file wins on procedure (how to ssh, lock, build, bench, profile).

Read order for a cold start:

1. This file (TL;DR + hard gates + quickstart).
2. VLLM_PARITY_FINAL.md (the closed record, every number cites its artifact).
3. .agents/vllm-parity-methodology.md (the methodology: bit-exact gating, profile-don't-assume, both-engine ground truth).
4. The patch-series README.md (~44 KB, canonical backend doc) and PAGED_BITEXACT_NOTE.md.

---

1. TL;DR STATE

• The investigation is CLOSED. Parity is not reachable on GB10 silicon; the residual is a hardware ceiling, not engineering debt.
• Prefill is a genuine floor at ~36% (MoE) / ~43% (dense) of vLLM. Prefill is not CUDA-graph-replayed, so these numbers are real, not measurement artifacts.
• Decode is near-parity: ~86% of vLLM's TRUE GPU-steady decode (924 vs 1078 t/s). The long-standing ~56% headline was a CUDA-graph measurement artifact (nsys without --cuda-graph-trace=node collapses each graph replay into one opaque launch). Decode is also ahead of vLLM at low concurrency (dense 116.7% at N=8) and uses 1.5-3x less memory, bit-exact per-path.
• The lever search was exhaustive: every attempt (prefill GEMM, GDN chunked scan, decode fusions, serving/scheduler) is recorded with its verdict and number so it is not re-run.
• The path to parity is different hardware: datacenter Blackwell (B200, HBM, native tcgen05 / CUTLASS FP4). Do NOT reopen GB10 kernels. Re-run the methodology on the new silicon, where vLLM's GB10-losing FLA/Marlin kernels invert.

---

2. THE HARD GATES YOU MUST NOT VIOLATE

These are non-negotiable. Violating any of them invalidates the result or the contribution.

2.1 The per-path greedy-md5 bit-exact gate (sacred)

The gate is per-path: paged vs non-paged attention legitimately produce different (equivalent) FP-reduction orders. Each path is gated against its own reference, validated benign by KL-divergence to the f16 reference. Canonical greedy md5s:

Path	Model	Canonical md5
non-paged	MoE q36-35b-a3b-nvfp4	07db32c2bcb78d17a43ed18bc22705cd
paged	MoE q36-35b-a3b-nvfp4	8cb0ce23777bf55f92f63d0292c756b0
non-paged	dense q36-27b-nvfp4	5951a5b4d624ce891e22ab5fca9bc439
paged	dense q36-27b-nvfp4	5951a5b4d624ce891e22ab5fca9bc439 (bit-exact to non-paged)

• Compare paged-to-paged only. Future paged-MoE regressions compare to 8cb0ce23, NOT 07db32c2.
• Why paged-MoE differs (benign, KL-validated): llama-perplexity --kl-divergence on the MoE GGUF (16 chunks, f16 base PPL 7.3734) shows non-paged-vs-f16 KLD 0.136597 and paged-vs-f16 KLD 0.136000, i.e. paged does NOT diverge from f16 ground truth more than non-paged does. Paged and non-paged are two equivalent FP-reorderings of the same 4-bit model. This holds on the 0028 baseline and with LLAMA_MOE_FORCE_GRAPHS/0029 on or off, so it is a property of the paged path, not any one lever.
• Every bit-exact patch is gated two ways: greedy md5 (per path) AND test-backend-ops vs the CPU oracle for every touched op.

2.2 The KL-gate for opt-in lossy paths

Any path that is NOT byte-identical (e.g. 0033 dequant-bf16, the 0034/0035 large-M FP paths, FP8-KV) ships default-off and is gated by a KL-divergence band: it requires KLD(new||f16) <= KLD(FP4-MMQ||f16) and PPL within the established band. Lossy levers never ship default-on.

2.3 In-backend A/B is the only proof (hard methodology rule)

A lever compiled into the binary is NOT isolated by a runtime flag alone. It needs a separately-built in-backend A/B. Precedents that burned this in: 0031 chunking math was correct yet -22% in-backend; 0034 had a standalone PoC win that did not hold in-backend.

2.4 Contribution / commit gates (LocalAI policy)

• DCO sign-off is human-only: do not add an AI Signed-off-by trailer.
• AI attribution via Assisted-by: trailer: Assisted-by: Codex:gpt-5.
• NEVER add Co-Authored-By: (AI) trailers and never add an AI Signed-off-by.
• No em-dashes anywhere in output (use -, :, parentheses, or rephrase).
• Ask before every git push. Prior approval does not carry over.

2.5 Fork-first is MANDATORY (the fork is canonical)

• The canonical source of truth is the fork branch mudler/llama.cpp:localai-paged (= pin commit + paged patch commits in order). It is canonical for ALL paged-backend kernel/patch work. The shipped patches/paged/*.patch series is a derivative: the fork is the source.
• Always update the fork FIRST, in this exact order: (1) commit the change on the localai-paged branch and push it, then (2) regenerate the LocalAI series (backend/cpp/llama-cpp-localai-paged/patches/paged/) from the fork via git format-patch (one patch per fork commit, source-only, never touching a *.md/dev-doc), so the series stays a 1:1, drift-free mirror of the branch. No hand-export.
• NEVER edit the LocalAI patches/paged/*.patch files directly, and NEVER add a patch to the series with no corresponding fork-branch commit. They are generated output, not source.
• The fork branch is also where the build and the per-path bit-exact md5 gate actually run, so it is the only place a change is truly validated. A patch that lives only in the LocalAI series has never been built or gated.
• Mirror invariant (verify by tree hash): applying the full on-disk series on the pin must reproduce the fork branch tree byte-for-byte. The series has intentional gaps (missing 0005, 0026, 0027, 0032, 0036-0039, 0045), so the patch count is not the max number; what must hold is the tree-hash equality, not the count. Current verified state: fork HEAD fb9402661 is mirrored by worktree patch 0055-feat-server-trace-speculative-batch-shapes.patch; applying all 46 patch files on 0ed235ea2c17a19fc8238668653946721ed136fd produces tree 5bdbf8ea3d750fe6fa1f85175fd6357d36222edb, exactly matching the fork.

2.6 Bench hygiene gates

• NEVER set LLAMA_MAX_BATCH_TOKENS in benches (the harness explicitly logs "NO LLAMA_MAX_BATCH_TOKENS").
• Do not set GDN_TC, GDN_CHUNK_MIN, or LLAMA_PAGED_DECODE_STABLE in parity benches. Production defaults are compiled in: GDN M5 on (GDN_TC=5, GDN_CHUNK_MIN=64), S1 decode-graph on, S3 off.
• Decode profiling MUST use nsys --cuda-graph-trace=node (see section 3.4). This is a gate, not a suggestion.

---

3. OPERATIONAL QUICKSTART (copy-pasteable)

3.0 Host

ssh dgx.casa        # resolves to hostname promaxgb10-4ad8; GPU = NVIDIA GB10 (unified LPDDR5x, ~273 GB/s, the bandwidth floor)

nvidia-smi reports memory as [N/A] (unified memory). CUDA 13 / sm_121.

3.1 GPU lock protocol (~/gpu_bench_lock) - TWO conventions, reconcile carefully

There are two conventions in flight:

• Old harnesses (combined_definitive.sh, fuse_validate.sh, fuse_profile.sh) treat it as an empty mutex dir: mkdir ~/gpu_bench_lock to acquire, rmdir to release.
• Newer harnesses (fp4norm_profile.sh) use an owner-file convention: mkdir -p ~/gpu_bench_lock then echo "$ME $(date +%s)" > ~/gpu_bench_lock/owner. They poll until nvidia-smi --query-compute-apps=pid count is 0 AND owner is FREE*/absent for 2 consecutive checks, and clear a stale ~/gpu_bench_lock/release file. Release writes FREE released-by-... $(date +%s) to owner (it does NOT remove the dir).

Because the dir now permanently contains an owner file, release with rm -rf ~/gpu_bench_lock, NOT rmdir (rmdir fails on the non-empty dir). Recommended procedure for a future agent:

1. Read ~/gpu_bench_lock/owner. FREE*/absent + 0 compute-apps means free.
2. Acquire via mkdir -p ~/gpu_bench_lock + write owner.
3. Release by writing FREE ... to owner (or rm -rf ~/gpu_bench_lock).

A separate 0-byte ~/bench/gpu.lock is legacy/unrelated - ignore.

Always gate on ALL THREE before benching or building on DGX: nvidia-smi --query-compute-apps=pid count == 0, owner FREE, and docker ps shows no running containers. In particular, do not start work while a local-ai-worker container is running. Concurrent jobs share this GPU: an offline-repack Marlin workflow, an ~/.cache/autoresearch-quant/ quant pipeline (this is the llama-imatrix class of job), finetune trees, and LocalAI worker containers. The canonical harnesses poll for GPU-idle up to 2h.

3.2 Build (long; run detached + poll)

• Mainline / canonical grpc-server + binaries: CUDA arch 121 (-DCMAKE_CUDA_ARCHITECTURES=121). Runtime banner shows ARCHS = 1210 | BLACKWELL_NATIVE_FP4 = 1.
• FP4-MMA / tensor-core experimental kernels: the accelerated 121a gencode (arch=compute_121a,code=[compute_121a,sm_121a]). The a suffix unlocks tcgen05 / native FP4-MMA intrinsics. 121a lives ONLY in the DGX experimental build scripts (~/gdn_cc.sh standalone nvcc, ~/gdn_bv_build.sh -DCMAKE_CUDA_ARCHITECTURES=121a, ~/paged-build.sh --build-arg CUDA_DOCKER_ARCH=121a), not in the worktree build files. Supply it at build time via CMAKE_CUDA_ARCHITECTURES / CUDA_DOCKER_ARCH.
• Long builds: run detached and poll for a marker. Pattern: nohup ... > build.log 2>&1 & then poll for a .DONE/.done file. Do NOT block a foreground shell.

Built binaries live at dgx:~/llama-paged-dev/build-cuda/bin/ (llama-server, llama-batched-bench, llama-completion; thin ~70 KB dynamic wrappers).

3.3 The standard bench env + commands

cd /home/mudler/llama-paged-dev/build-cuda/bin
L="LLAMA_KV_PAGED=1 LLAMA_MOE_FORCE_GRAPHS=1 GGML_NO_BACKTRACE=1"   # GGML_NO_BACKTRACE is log-hygiene, not a lever
MOE=/home/mudler/bench/q36-35b-a3b-nvfp4.gguf       # arch qwen35moe, ~22.2 GiB
DENSE=/home/mudler/bench/q36-27b-nvfp4.gguf         # arch qwen35,    ~17.5 GiB

# (1) Bit-exact / coherence gate. stdin MUST be /dev/null or it hangs in conv mode.
env $L ./llama-completion -m "$MOE" -ngl 99 -fa on -c 4096 --temp 0 --seed 1 -n 48 -no-cnv \
    -p "The capital of France is" </dev/null | md5sum
# The PAGED_BITEXACT_NOTE gate command uses the chat-template path (NO -no-cnv):
#   ./llama-completion -m MODEL -ngl 99 -fa on -p "The capital of France is" -n 48 --temp 0 --seed 1
# (compare to the canonical md5 for that model+path; paged-to-paged only)

# (2) PREFILL bench (S_PP from llama-batched-bench)
env $L ./llama-batched-bench -m "$MOE" -c 131072 -b 2048 -ub 512 -ngl 99 -fa on \
    -npp 512,2048 -ntg 4 -npl 32

# (3) SERVING bench: one --parallel 256 server, then drive with h2h_cli3.py
env $L nohup ./llama-server -m "$MOE" -c 262144 --parallel 256 -b 2048 -ub 512 \
    -ngl 99 -fa on --host 127.0.0.1 --port 8090 --no-webui >/home/mudler/bench/paged_server.log 2>&1 &
# poll http://127.0.0.1:8090/health for '"ok"', then:
python3 /home/mudler/bench/h2h_cli3.py   # OpenAI /v1/completions, ignore_eos, fresh-nonce, ptok128 gen128, NPL sweep 8/32/128/256

vLLM side (for both-engine parity): ~/vllm-bench/bin/vllm (version 0.23.0), served gpu-util 0.85 max-model-len 4096 max-num-seqs 256 tp1, models ~/bench/q36-35b-a3b-nvfp4-vllm/ and ~/bench/q36-27b-nvfp4-vllm/.

Current-stack serving snapshots use backend/cpp/llama-cpp-localai-paged/paged-current-serving-snapshot.sh. It targets the clean ~/llama-phase6-source mirror, checks docker/local-ai-worker/GPU-idle state, uses the owner-file lock, runs pre/post inference gates, then compares paged and vLLM with the same h2h client. The older dgx:~/bench/combined_definitive.sh is historical: do not reuse it without first porting away from stale ~/llama-paged-dev paths and old lock assumptions.

3.4 THE DECODE-PROFILING RULE (this trap caused 4 wrong analyses)

Decode runs as a replayed CUDA graph. nsys without --cuda-graph-trace=node collapses each graph replay into ONE opaque launch, so every per-kernel attribution becomes an artifact. This is exactly what made the old "paged 159 us/tok, GPU ~16% busy, host-bound, 5.4x more GPU-efficient" story wrong, and produced the wrong ~56% headline.

Mandatory method for any decode profile:

• Use nsys --cuda-graph-trace=node.
• Decompose with the difference method: per-token cost = (ntg=64 profile) - (ntg=16 profile).

Under the correct method, paged decode at npl=256 is 99% GPU-busy (1.4% idle), NOT host-bound - the opposite of the collapsed-graph reading. The clean graph-node-traced profiles are at ~/highN_prof2/*.nsys-rep (paged, npl=256) and ~/highN_vllm/*.nsys-rep (vLLM), captured 2026-06-30. They supersede every earlier decode decomposition.

3.5 Models + artifacts (all on DGX)

GGUF (paged): ~/bench/q36-35b-a3b-nvfp4.gguf (MoE, qwen35moe), ~/bench/q36-27b-nvfp4.gguf (dense, qwen35). vLLM safetensors: ~/bench/q36-35b-a3b-nvfp4-vllm/ (has hf_quant_config.json confirming MIXED_PRECISION / FP8-proj), ~/bench/q36-27b-nvfp4-vllm/. Authoritative run: ~/bench/COMBINED_DEFINITIVE.txt (+ .log, .done, combined_definitive.sh, per-engine COMBINED_*_server.log). A/B dirs: ~/bench/marlin_gate/, ~/bench/gdn_p1_ab/. NOTE: the *_RESULTS*/*_MAP* docs live only in the worktree docs/, not on the DGX.

---

4. THE COMPLETE LEVER MAP (do NOT re-run the rejected ones)

Verdicts and numbers are from VLLM_PARITY_FINAL.md + the cited artifacts. "BE" = greedy-md5 bit-exact; "KL-benign" = lossy path inside the KL band.

4.1 Prefill weight-GEMM track - WHOLE TRACK REJECTED (FP4-MMQ is optimal on GB10)

Decisive surprise: on sm_121 vLLM itself does NOT run native FP4 - it runs Marlin W4A16 (FP4 dequant->bf16 in-register + bf16 GEMM) for experts and FP8 projections, capped at ~half FP4 peak, because native CUTLASS NVFP4 grouped-GEMM is broken on consumer Blackwell (TMA-WS init failure, CUTLASS #3096; no tcgen05/TMEM). So MMQ's native FP4 is already structurally competitive here.

Lever	What	Verdict	Key number
0033 dequant->bf16 cuBLAS	route large-M NVFP4 dense GEMM to dequant->bf16 cuBLAS	REJECTED, ships default-off	dense S_PP -49%/-42%/-29% at M=512/1024/2048; BE + KL-better
dense-cuBLAS reroute (full)	same across dense+MoE prefill	REJECTED	-31% to -62% band
0034 native FP4-MMA W4A4	Blackwell mxf4nvf4 OMMA large-M	REJECTED in-backend	PoC 103 TFLOP/s (57.7% FP4 peak, NMSE=0) but win did not hold in-backend
0035 W4A16-Marlin grouped MoE	FP4->bf16 in-register + bf16 mma, zero act-quant tax	REJECTED (perf)	correct + KL-benign-and-better but -39% S_PP vs MMQ
0045/0046 offline-repack / vLLM-verbatim Marlin	repack to Marlin layout; port vLLM kernel verbatim	REJECTED	verbatim correct but -39%; offline-repack same bf16-peak ceiling, no win

Why it loses: bf16 TC peak on GB10 is ~half FP4 peak, so any dequant->bf16 kernel caps at ~half FP4-MMQ; the dequant write is an un-amortized weight-sized memory pass (~8x the FP4-read traffic). The GEMM bucket is not winnable on GB10 with available kernels.

4.2 Prefill GDN chunked-scan track - M5 tf32 C=16 is the SHIPPED winner

GDN is the #1 prefill-gap contributor (+59.2 us/tok, ~30%). vLLM's FLA chunk_gated_delta_rule runs the same math at 36.5 vs paged 95.7 us/tok = 2.62x via tensor-core intra-chunk Gram products.

Lever	What	Verdict	Key number
0031 scalar-serial chunked scan	FLA-style scalar/serial (GDN_TC=0)	superseded	correct but ~22% slower at forced C=16
0047 / M5 tf32 tensor-core scan	tf32 m16n8k8 mma form-T solve, f32-only	SHIPPED default-on under paged	MoE prefill +3.5% @npp512, +17.7% @npp2048; decode unchanged; BE-benign
bf16 CONFIG-C (M8)	bf16 Kc/Qc + 2 C*C scratch, C->64	REJECTED (not in f32 series)	confirmed geometry then dropped
bf16-C16	bf16 Gram at C=16	REJECTED	no win; bf16 mantissa unsafe on state-coupled products
BV block-occupancy A/B (tf32)	raise blocks/SM	REJECTED (occupancy NOT the bound)	1844 vs 1814 S_PP (-1.04%, within noise)
bf16-C64	bf16 Gram at C=64	REJECTED	-18.75%; O(C^2) intra-chunk + serial recurrence dominates
Phase 10 C32 slab M5	C=32, two dv_tile=64 slabs, default-off GDN_C32_SLAB=1	REJECTED	md5-clean after tail-row zeroing, but slower: MoE 2048 2430.32 -> 2054.86; dense 2048 1019.25 -> 903.73
Phase 11 QS-early M5	move QS = Qc * S0 earlier, default-off GDN_M5_QS_EARLY=1	REJECTED	md5-clean, but slightly slower: MoE 2048 2441.54 -> 2420.26; dense 2048 1021.06 -> 1015.77
Phase 12 shared-A/Ai cost model	f32 Ai scratch shared across two C32 value slabs	GO to one default-off prototype	BT32 f32 scratch at npp2048,npl32: MoE 256 MiB / 768 MiB Ai traffic; dense 384 MiB / 1152 MiB Ai traffic
Phase 13 Global-Ai32	precompute f32 Ai once, consume from two C32 dv_tile=64 slabs	REJECTED	md5-clean, but slower: MoE 2048 2425.10 -> 2097.76; dense 2048 1016.14 -> 918.19

Why not occupancy/dtype: the cost is the O(C^2) intra-chunk triangular A-inverse solve + the strictly-serial inter-chunk recurrence, with C forced to 16 by GB10's 99 KB dynamic-smem cap (the 128x128 f32 state alone is 64 KB). M5 captures the tractable TC part; it does not fully close 2.62x because vLLM's FLA blocked-solve is a more complete TC implementation.

Phase 13 closes the caveat: the default-off GDN_GLOBAL_AI32=1 prototype was correctness-clean but slower. Stop GDN kernel work on GB10 instead of iterating into f16 Ai or more local reorders.

4.3 Decode / fusion levers - all REJECTED (near-parity already at ~86% true GPU-steady)

Lever	What	Verdict	Key number
act-quant folded into ggml MMQ	inline y-quant in MoE expert MMQ	REJECTED	-79.4%; ggml MMQ re-quantizes y per weight-row-tile x stream-k split, no TC for inline quant
norm+quant+silu fusion	one launch (vLLM Triton kernel)	REJECTED (infeasible)	ggml_cuda_can_fuse cannot express it: FP4 quant is a mul_mat-internal prologue, silu separated from norm by 2 GEMMs + router
Q8_0 / FP8 projection	quantize bf16 GDN/attn projections	REJECTED (regime error)	vLLM DOES use FP8 proj, but at N>=128 proj is only ~12% of stream, closes <=6%
NVFP4 the projections	drop proj to NVFP4	REJECTED	KL-fail, ~+6% PPL; vLLM keeps SAME bf16/FP8 proj, never NVFP4
W4A16-Marlin MoE decode	Marlin grouped expert GEMM at decode	REJECTED	BW-floored wash, ~5% slower
bf16-tau per-head SSM (0026)	per-head bf16 tau on SSM decode	DROPPED	flat 780.6 vs 780.0 t/s; earlier "+12%" subsumed by 0028/0029
D3 FA-split / D4 GDN-width-adaptive	older off-critical-path levers	SUPERSEDED reasoning	were rejected via the debunked "5.4x/host-bound" reading; under HNP the GDN scan IS critical path (51%) but is the shared BW floor where paged leads (83% vs 79%), so still not a win

Dense decode is AHEAD at low N (116.7% @ N=8) - the one operating point where paged is unambiguously faster.

4.4 Serving / engine levers - host loop and scheduler CLOSED

Lever	What	Verdict	Key number
0040 / S1 paged decode-graph reuse	can_reuse keyed on bucketed block-table dims	SHIPPED default-on	serving reuse 0% -> 72.2% (with S3); static 0% -> 95.5%
0041 / S3 decode-shape-stable scheduling (LLAMA_PAGED_DECODE_STABLE)	keep prefill out of decode steps	SHIPPED default-OFF (opt-in)	recovers the ~17 pt graph-reuse overhead at a TTFT cost; default-on regressed real serving (2.5x worse TTFT, 20-29% lower e2e throughput)
0043 / D1 full-step MoE decode CUDA graph	graph whole decode step incl. grouped-MMQ MoE dispatch	SHIPPED default-on	+2.6% (npl128) to +5-13% (npl32); D1 premise "host-sync on MoE readback" REFUTED (sync count identical 1457 on/off)
S2 double-buffer set_inputs	overlap host input build with GPU	DROPPED	set_inputs ~0.05 ms/step, nothing to recover
whole-step graph / host loop	host loop as serving residual	CLOSED (~0-1%)	reuse 0% (757.6) == S1+S3 72% (763.3); hostproc only ~4-8% of step wall
padded / fixed-slot decode	pad decode width to --parallel for ~100% reuse	REJECTED (built, GPU-tested, commit b028c81e)	inert (BE) but regresses everywhere; N=8 burst 28.16->6.05 tok/s/seq; serving decode is GPU-compute-bound, dummy-row compute > reuse recovered
speculative decode (MTP)	draft + verify	REJECTED for current GB10 serving	Phase 14 safety passed, but Phase 15 serving A/B regressed hard: n128 decode agg 662.4 -> 138.5 tok/s; likely graph/batch-shape disruption (graphs reused 361 -> 1)

4.5 SHIPPED WINS (all BE / KL-benign) - keep these, do not regress

• FP4-MMQ MoE/dense GEMM (native Blackwell FP4-MMA at the FP4 weight-BW floor; reason 4.1 stays default-off).
• M5 tf32 tensor-core chunked GDN prefill (patch 0047), default-on under LLAMA_KV_PAGED (GDN_TC=5 + GDN_CHUNK_MIN=64).
• 0042 fused residual-add + RMSNorm + weight-mul (dense S_PP +0.5%, BE).
• 0044 fused GatedRMSNorm + SiLU gate-mul (672 -> 336 launches @npp512; dense +1.1%, MoE +0.9%, test-backend-ops 12979/12979).
• 0046 GDN-prefill geometry gate (gates 0022's decode retune by scan length; recovers +7.2% dense prefill, keeps the decode win, BE).
• SSM decode-fusion stack (0018-0022, 0028): in-place state (+23.5%/+18.9%), fused gather (+37.8%/+35.3%), o_proj reshape (+31.7%/+23.3%), conv in-place (+3.2%/+3.5%), occupancy retune (+11.1%/+8.3%) = the 2.26x / 2.46x over stock decode multiplier.
• Serving host loop closed (0040 S1, 0043 D1).
• The memory advantage (1.5-3x lower VRAM, NVFP4-resident, no persistent bf16 dequant copies).
• Low-N decode lead (dense 116.7% @ N=8). Bit-exact output per-path through the whole series.

4.6 REMAINING / unattempted levers + EV

• Multi-week persistent-Marlin decode kernel (vLLM's fused-Marlin MoE persistent-tiling + Triton elementwise): the only path to the residual ~14 pt GPU-steady decode gap. Low-EV: decode-only ~4-14%, our own ggml Marlin port already lost -19.6%, needs mature tiling + multi-stream overlap (hard inside a single-stream CUDA graph), GB10-uncertain, and cannot lift the prefill floor. Not a free bit-exact lever.
• Datacenter-Blackwell pivot (B200, ~8 TB/s HBM, native tcgen05/CUTLASS FP4, TMEM): lifts the LPDDR5x GDN bandwidth floor ~30x and restores exactly the vLLM advantages that lose on GB10. This is the documented path to parity. Re-run the methodology on the new silicon, do not reopen GB10 levers.

The VLLM_PARITY_LEVER_MAP.md "pursue list" (A1-A7/B1-B7/C1: graph-safe ragged grouped FP4-MMA MoE kernel, FP8 paged KV, MTP spec-decode, etc.) is the earlier working brainstorm written before the final profiling. VLLM_PARITY_FINAL.md is the authoritative supersession; treat those buckets as rejected / infeasible / different-hardware unless re-validated on new silicon.

Phase 14 re-validated the MTP bucket as safe, then Phase 15 rejected it as a current GB10 serving-throughput lever. Do not enable it by default and do not keep tuning draft length blindly. The only plausible follow-up is a graph-reuse and speculative verification batch-shape profile with nsys --cuda-graph-trace=node. Phase 16 ran that profile and supported the root cause: small-shape baseline reused graphs (graphs reused = 62) while MTP did not (graphs reused = 1) and did ~2.3x more GPU kernel work. The fixed safety gates stayed green before and after the failed serving A/B: MoE md5 8cb0ce23777bf55f92f63d0292c756b0, dense md5 5951a5b4d624ce891e22ab5fca9bc439, and MUL_MAT_ID 806/806.

Phase 17 source inspection found no tiny additive graph-reuse fix. MTP verification rows are real target decode/output rows (K + 1 per speculative slot), so fake padding would touch KV, positions, logits, MTP nextn state, and rollback semantics. If reopened, start with a server-only shape counter around server_slot::handle_last_sampled_token(). Only then consider an opt-in group/defer-by-draft-length scheduler experiment, with TTFT/throughput and md5/op gates as kill criteria.

Phase 18 added the server-only shape trace as patch 0055. Set LLAMA_SPEC_SHAPE_TRACE=1 to log kind=decode rows and MTP kind=verify K + 1 row/output shapes from server_slot::handle_last_sampled_token(). This is default-off instrumentation only. DGX green check after the patch saw MTP verify shapes vary (rows=4, then rows=3) on a tiny request, while the env-unset run emitted no spec shape: lines. Canonical post-patch gates passed: MoE 8cb0ce23777bf55f92f63d0292c756b0, dense 5951a5b4d624ce891e22ab5fca9bc439, and MUL_MAT_ID 806/806. Artifacts: /home/mudler/bench/phase18_mtp_shape_trace_green and /home/mudler/bench/phase18_mtp_shape_trace_green/gate_after.

Next MTP step, if any: trace real serving shape entropy first. Do not implement a scheduler change until the trace shows repeatable draft-length buckets worth grouping. Any scheduler experiment must be opt-in/default-off and killed by TTFT/throughput regression, graph-reuse failure, md5/op drift, or MTP rollback/prefix gate failure.

Phase 19 ran that trace-only serving measurement and rejected the scheduler shortcut. Artifact: /home/mudler/bench/phase19_mtp_shape_entropy/20260701_045534. Pre/post gates passed with canonical MoE md5 8cb0ce23777bf55f92f63d0292c756b0, dense md5 5951a5b4d624ce891e22ab5fca9bc439, and MUL_MAT_ID 806/806.

Serving result:

n	baseline decode_agg	MTP decode_agg	MTP / baseline	baseline TTFT ms	MTP TTFT ms
8	245.0	95.7	39.1%	1147.2	1633.4
32	409.2	110.0	26.9%	2710.0	4471.5
128	697.2	154.0	22.1%	7601.5	20310.4

Shape result: draft=3 already accounts for 96.2-96.9% of verify slots, so group/defer-by-draft has little to recover. Full in-flight steps already mostly use all-draft=3 vectors; the remaining churn is active-slot/tail churn plus the real K + 1 verification-row expansion. Do not build a Phase 20 scheduler experiment on this evidence. Future MTP work would need a deeper target-verify graph/state design, not another small server scheduling shortcut.

Phase 20 refreshed the current-stack MoE serving snapshot against vLLM using the clean ~/llama-phase6-source mirror (f2521ab12) rather than the stale llama-paged-dev benchmark tree. Artifact: /home/mudler/bench/phase20_current_snapshot/20260701_050621. Pre/post gates passed with canonical MoE md5 8cb0ce23777bf55f92f63d0292c756b0, dense md5 5951a5b4d624ce891e22ab5fca9bc439, and MUL_MAT_ID 806/806.

Current MoE serving snapshot (PTOK=128, GEN=64):

n	paged decode_agg	vLLM decode_agg	paged/vLLM decode	paged agg	vLLM agg	paged/vLLM agg
8	220.8	290.5	76.0%	164.8	245.5	67.1%
32	411.1	594.7	69.1%	252.1	456.0	55.3%
128	670.0	1022.7	65.5%	322.4	662.4	48.7%

TTFT remains the clearest user-visible gap: paged is 2.88x/3.36x/3.11x slower than vLLM at n8/n32/n128, and paged prefill_tps is roughly one-third of vLLM. This keeps the GB10 shortcut closure intact: do not reopen MTP or small scheduler work. The credible next parity path is a datacenter-Blackwell rerun or a larger fused-kernel project outside this low-conflict patch stack.

Phase 21 added a reusable current-stack serving harness: backend/cpp/llama-cpp-localai-paged/paged-current-serving-snapshot.sh. It defaults to ~/llama-phase6-source, validates docker/local-ai-worker/GPU idle state, uses the owner-file lock, runs pre/post inference gates, compares paged and vLLM with h2h, and writes ratio summaries. DGX dry run passed at /home/mudler/bench/phase21_harness_dryrun/20260701_051757.

Use this harness for future current-stack GB10 snapshots. Do not reuse ~/bench/combined_definitive.sh unless it is first ported away from stale ~/llama-paged-dev paths and old lock assumptions.

Phase 22 re-verified the patch-series mirror invariant after patch 0055: applying every LocalAI patches/paged/0*.patch with strict git apply on top of Makefile pin 0ed235ea2c17a19fc8238668653946721ed136fd produced tree 5bdbf8ea3d750fe6fa1f85175fd6357d36222edb, exactly matching fork branch localai-paged HEAD fb9402661 feat(server): trace speculative batch shapes.

---

5. METHODOLOGY LESSONS (so you do not repeat the mistakes)

1. Profile, don't assume. The analysts were wrong 4 times. Every one was caught only by an in-backend A/B or a corrected profile:
   • GDN-scalar grep (assumed the scan was scalar/serial from reading source) - wrong, retired by the tensor-core port.
   • dense-cuBLAS reroute (assumed dequant->bf16 would win) - wrong, -31% to -62%.
   • occupancy (assumed blocks/SM was the GDN bound) - wrong, 1844 vs 1814 within noise.
   • projection-regime (assumed FP8/NVFP4 projections were a big lever) - wrong, projections are ~12% of the decode stream at high N. In-backend A/B is the only truth. A standalone PoC win (0034) is not a result.
2. Per-kernel us/tok overstates end-to-end S_PP/S_TG. A kernel that is X% faster in isolation does not move throughput X%; always confirm against the end-to-end batched-bench / serving number.
3. The CUDA-graph-trace decode artifact (the big one). Decode is a replayed graph; nsys without --cuda-graph-trace=node collapses it and lies. This single trap produced the wrong "host-bound / 159 us/tok / 56%" story across multiple analyses. Always graph-node-trace + difference method (section 3.4).
4. Beware GPU contention skewing absolutes. The box runs concurrent quant/repack/finetune jobs. Gate on idle GPU + free lock; prefer the same-session both-engine harness so both numbers move together.
5. The vLLM server number is inflated ~8 pt vs its true GPU-steady. vLLM's chunked-prefill-overlap inflates its own server-measured decode window (1177 server vs 1078 true GPU-steady). Compare GPU-steady to GPU-steady, or you will chase a phantom gap. The reconciliation chain that must sum: vLLM server 1177 (100%) -> vLLM true GPU-steady 1078 (92%) -> llama GPU-steady 924 (78.5% of 1177, = 86% of 1078) -> llama server 718 (60.7%, the S3-recoverable serving overhead).

---

6. THE THREE FORWARD DIRECTIONS

(a) Close / ship the record (lowest effort, do this first)

The investigation is closed for GB10 shortcuts, and the closeout chores below are now done:

• patch 0044 is tracked in the LocalAI series;
• the Makefile pin 0ed235ea2c17a19fc8238668653946721ed136fd is the authoritative paged pin;
• Phase 20 re-ran the current-stack serving snapshot on the clean mirror;
• Phase 22 re-verified the patch-series mirror invariant after 0055.

For future release checks, run paged-inference-gates.sh and paged-current-serving-snapshot.sh from the LocalAI backend tree.

(b) Datacenter-Blackwell pivot (THE real parity path)

The thesis: every vLLM advantage that wins on GB10 is a kernel that is broken or capped on consumer Blackwell and inverts on datacenter Blackwell (B200): FLA blocked-solve GDN, Marlin/CUTLASS grouped FP4, HBM-tuned full-cudagraph decode, native tcgen05/TMEM. ~8 TB/s HBM lifts the LPDDR5x GDN bandwidth floor ~30x. Concrete first steps:

1. Acquire a B200 (or equivalent HBM tcgen05 part). Reproduce the both-engine same-session harness there (combined_definitive.sh discipline): build the stock and paged binaries, build vLLM 0.23.0+, run MoE + dense prefill + serving for both engines.
2. Re-measure the FP4 path: on B200, native CUTLASS NVFP4 grouped-GEMM should work (the CUTLASS #3096 / TMA-WS failure is consumer-Blackwell-specific). Confirm whether vLLM now runs native FP4 instead of Marlin W4A16. If so, the 4.1 GEMM track must be re-evaluated from scratch (it was rejected on a GB10-specific ceiling).
3. Re-take the decode profile with --cuda-graph-trace=node; the GDN scan that floors at 273 GB/s on GB10 should no longer dominate at HBM bandwidth - re-derive the per-token decomposition before choosing any lever.

(c) Multi-week persistent-Marlin decode kernel (decode-only, low-EV, CANNOT reach parity)

Only pursue if (a)+(b) are not options and someone explicitly wants the residual decode gap closed on GB10. It targets the ~14 pt GPU-steady decode gap (vLLM's fused-Marlin MoE persistent-tiling + single Triton elementwise). Concrete first steps:

1. Re-confirm the ceiling first: our own ggml Marlin port already lost -19.6% at decode (4.3), so the bar is "beat that and beat FP4-MMQ at the decode BW floor".
2. Prototype the persistent-tiling grouped-FP4 MoE kernel standalone, then prove it in-backend (a PoC win is not a result, per 0034). It must live inside a single-stream CUDA graph or bring its own multi-stream overlap.
3. Bound the upside honestly: this is decode-only ~4-14% and does nothing for the prefill floor (36-43%), so it does not reach parity. Record the verdict either way.

---

7. KEY FILE / ARTIFACT INDEX

Fork (canonical source of truth)

• Local canonical fork: /home/mudler/_git/llama.cpp, branch localai-paged, HEAD fb9402661291e0488a3e2bf2f3948ebcd18e18c9 ("trace speculative batch shapes", patch 0055).
• DGX current clean mirror/build tree: dgx:~/llama-phase6-source, HEAD f2521ab12 with the same tree as the local fork; this is what Phase 20 and the current snapshot harness use.
• Historical DGX dev tree: dgx:~/llama-paged-dev, branch paged, HEAD a7d439e8ce6990eb09721223c975da4e49d8d136 ("GDN CONFIG C (M8) - bf16 Kc/Qc"). It is an old experimental tree and must not be treated as canonical.

LocalAI worktree

• Path: /home/mudler/_git/LocalAI/.claude/worktrees/feat+paged-attention, branch worktree-feat+paged-attention (currently 246 ahead, 31 behind origin/master; recompute before reporting).
• Backend dir: backend/cpp/llama-cpp-localai-paged/ (Makefile thin wrapper, package.sh, run.sh, README.md ~44 KB canonical, docs/, patches/paged/).
• docs/: VLLM_PARITY_FINAL.md (authoritative record), VLLM_PARITY_LEVER_MAP.md (working brainstorm, profile-validated section), DECODE_SERVING_SCOPE.md, PREFILL_GEMM_SCOPE.md, PREFILL_GEMM_RESULTS.md, TENSORCORE_GDN_SCOPE.md, TENSORCORE_GDN_BUILD_PLAN.md, ACCELERATOR_PORTING_SCOPE.md, UPSTREAM_LAYER2_SCOPE.md, LOCALAI_LLAMACPP_BACKEND_PLAN.md, PAGED_BITEXACT_NOTE.md, PATCH_MAINTENANCE.md, final_benchmark.csv, paged-burst-bench.cpp, paged-reclaim-unit.cpp, 3 PNGs, and this PARITY_HANDOFF.md.
• patches/paged/: 46 .patch files spanning 0001-0055 with intentional gaps (missing 0005, 0026 [dropped ssm_bf16_tau], 0027, 0032, 0036-0039, 0045). Core paged-KV 0001-0012; decode-first scheduler 0013/0016; serving graph reuse 0040/0041; prefill fusions 0042/0044; SSM/GDN decode 0018-0022/0028; MoE NVFP4 quant 0023/0025/0043; FP4-MMA/Marlin scaffolds 0033/0034/0035 (default-off); GDN tensor-core prefill 0031 -> 0046 (geometry gate) -> 0047 (f32-only M5, default-on under paged KV); W4A16 packed metadata/shape/padding is 0048-0050; MoE safety tests are 0051-0053; MTP backend-sampling safety is 0054; speculative shape trace is 0055.

Bench artifacts (DGX)

• ~/bench/COMBINED_DEFINITIVE.txt (+ .log, .done, combined_definitive.sh, combined_definitive.out) - historical same-session both-engine run.
• ~/bench/phase20_current_snapshot/20260701_050621 - current clean-stack paged-vs-vLLM MoE serving snapshot.
• ~/bench/phase21_harness_dryrun/20260701_051757 - current snapshot harness dry-run artifact.
• Per-engine logs ~/bench/COMBINED_{paged,vllm}_{MOE,DENSE}_server.log; ~/bench/BENCHMARK_PROGRESS.md.
• Graph-node-traced high-N profiles: ~/highN_prof2/*.nsys-rep (paged npl=256), ~/highN_vllm/*.nsys-rep (vLLM), 2026-06-30.
• A/B dirs: ~/bench/marlin_gate/, ~/bench/gdn_p1_ab/.

Recent context commits

• 6edbb56b0 "docs(paged): definitive vLLM-parity final-state record (GB10, CLOSED)" - adds VLLM_PARITY_FINAL.md.
• baf102524 "docs(paged): correct decode-serving record to ~86% GPU-steady parity (graph-node-traced)" - the ~56% -> ~86% correction.
• bd100dd20 "fix(paged): repair the patch series, sync to the fork branch" - dropped dev-tree 0044/0045, added f32-only M5 as 0047.
• b028c81ed "docs(paged): record padded/fixed-slot decode shape as tested-and-rejected".

Discrepancies to flag / resolve (carried verbatim from the gather, including UNVERIFIED labels)

1. Pin prose reconciled in this worktree. Makefile line 52 LLAMA_VERSION?=0ed235ea2c17a19fc8238668653946721ed136fd is authoritative and matches the local fork merge-base. Hard rule: the paged pin must equal the stock llama-cpp pin (shared grpc-server.cpp); a bump to c299a92c once broke the grpc-server link despite being bit-exact and was reverted. Trust the Makefile when building.
2. Current fork/mirror are clean and verified. Local fork HEAD is fb9402661, DGX clean mirror HEAD is f2521ab12, and Phase 22 proved the LocalAI patch series tree equals the fork tree. The old llama-paged-dev tree is historical only.
3. Worktree patch series is tracked through 0055. The only current untracked path in this worktree is .claude/.
4. sm_121a is not in the worktree build files - it lives only in the DGX experimental build scripts (gdn_cc.sh, gdn_bv_build.sh, paged-build.sh); mainline uses arch 121. UNVERIFIED whether the shipped CI Dockerfile build path injects 121a for the FP4-MMA kernels (Dockerfile.llama-cpp-localai-paged does not hardcode a CUDA arch).
5. The 0921716... paged-MoE md5 open item. COMBINED_DEFINITIVE.txt records PAGED_GATE_MD5=0921716cd0582b5d15af8c362b811d00 for MoE, but a full doc/patch/git log -S grep of the worktree found no occurrence of 0921716... in any committed source; the committed canonical paged-MoE gate is 8cb0ce23. Treat this as unreconciled: the documented, KL-validated paged-MoE gate remains 8cb0ce23, and any paged-MoE divergence (including 0921716) must be KL-validated against the f16 reference before being accepted as benign, never on assertion alone. The 0921716 value is UNVERIFIED as a sanctioned gate; do not adopt it as canonical without re-running the KL gate. The dense run is symmetric: COMBINED_DEFINITIVE.txt records PAGED_GATE_MD5=ecfe924dee6c5622c149f419ff2a6481 for dense, which likewise differs from the canonical dense gate 5951a5b4. Both CDEF PAGED_GATE_MD5 values come from the combined_definitive.sh harness's own gate command, NOT the canonical bit-exact gate command in section 3.3, which is why they diverge from the committed 8cb0ce23 / 5951a5b4; neither is a sanctioned gate and both must be KL-validated before being treated as benign.

---

Status: investigation CLOSED. This handoff is procedure; VLLM_PARITY_FINAL.md is the record. The path to parity is datacenter Blackwell, not GB10 kernels.