serverAppConfigs/serverAppConfigs-docs/LLM/frank-medium-llamacpp-dflash2.md

type, title, description, tags, timestamp

type	title	description	tags	timestamp
configuration	frank-medium llama.cpp Runbook (DFlash2 drafter)	The ACTIVE frank-medium deployment since 2026-09-16 (model swapped 2026-09-19) — mainline llama.cpp on Dankserver serving orcarouter/Qwen3.8-27B-Uncensored-GGUF at Q8_0 with the z-lab DFlash2 Q4_K_M drafter (the model's own MTP head stays in the file, one SPEC_TYPE change away), as ONE layer-split process across both Radeon AI PRO R9700 cards at a balanced 32:33 split, a 327680-position KV pool (163840 per conversation over two slots), f16 KV plus a 16 GiB host-RAM prompt cache, vision via the BF16 mmproj.	frank-medium dankserver llama-cpp dflash2 dspark qwen35 gfx1201 r9700 layer-split	2026-09-19

frank-medium — llama.cpp (DFlash2 drafter)

The ACTIVE frank-medium deployment since 2026-09-16 (model swapped to orcarouter Q8_0 on 2026-09-19): mainline llama.cpp serving orcarouter/Qwen3.8-27B-Uncensored-GGUF at Q8_0 with the z-lab DFlash2 Q4_K_M block-diffusion drafter, as one process across both Radeon AI PRO R9700 cards, layer-split at a balanced 32:33. The predecessor model (ajgazin Swift-Qwen3.8-27B UD-Q8_K_XL, 35:30 skew, 393216 pool) is the calibration history in the sizing tables below. DSpark was tried in the drafter's place on 2026-09-16 and reverted — see The DSpark excursion. It replaced the vllm-radiance MXFP4 replica pair + routing proxy (frank-medium-radiance-mxfp4, the record and the PCIe-atomics evidence) in the same day the pair went live. It is also the box's only LLM runtime: the replicas, the proxy and their unit files are retired.

Why one process and layer split

The endpoint's whole history is one constraint: the two cards cannot take part in an RCCL collective. RCCL's kernels declare a hidden hostcall buffer, and ROCr refuses to supply it to a device without PCIe atomics; the second R9700 hangs off the B650 chipset (AtomicOpsCap: Routing-, read-only), so every collective that touches it dies with hipErrorIllegalState. That rules out tensor parallelism — and it rules out vLLM's pipeline parallelism too, whose stage handoff is a send/recv of the hidden state on the device backend (vllm/distributed/parallel_state.py), the very primitive that fails. The full matrix is in Dual-R9700-Cooperation.

llama.cpp's --split-mode layer needs no collective at all: layers are distributed across the cards in order and the boundary activation is a plain per-backend copy, once per forward pass. Measured on this host with the predecessor runtime (llama-bench, pp4096 / tg128, f16 KV, flash-attn on):

Split	pp4096	tg128
one R9700 (-sm none)	1160.9	19.42
two cards, layer (-sm layer)	1991.2	18.76
two cards, tensor (-sm tensor, no RCCL: butterfly fallback)	271.8	26.26

Tensor split buys ~40 % decode and loses 7× prefill. The workload here is prefill-heavy (agentic, long prompts, reasoning traces), so layer split is the shape — prefill nearly doubles and one model serves from both cards.

Hardware

Property	Value
GPUs	2x AMD Radeon AI PRO R9700 (DID 0x7551), gfx1201 (Navi 48, RDNA4)
VRAM	34,208,743,424 B (31.86 GiB) each, as rocm-smi reports it
Link	PCIe, no XGMI; card #2 behind the chipset, no atomic routing
iGPU	Raphael (gfx1036) enumerates as HIP device 2 — hidden via HIP_VISIBLE_DEVICES=0,1
CPU / RAM	Ryzen 7 7700X (8c/16t), 60 GiB RAM, 63 GiB swap

Runtime

Mainline llama.cpp at fb27a525 (2026-09-16), built for gfx1201 only by scripts/build-llama-rdna4.sh against a private ROCm 7.14.0a20260624 toolchain (TheRock nightly) that is bundled next to the binary — RPATH=$ORIGIN/../lib, no LD_LIBRARY_PATH, nothing from the host's ROCm 7.1 packages.

Two features pin that ref, and neither is in the ROCmFPX fork the endpoint used earlier the same day (its tip, c49ebdb, predates both):

• DFlash2 (ggml-org/llama.cpp#27342): grouped dynamic depthwise convolution plus the candidate selector, detected from the draft GGUF (selector_top_k > 0) and served through the ordinary --spec-type draft-dflash path.
• qwen35 MTP: the draft-mtp implementation covers this architecture's single trained head.

Build with:

bash scripts/build-llama-rdna4.sh     # ~25 min on this box, JOBS=$(nproc)

Bundling trap, measured. rocBLAS and hipBLASLt look for their device code at a fixed depth under the binary's lib dir (../lib/rocblas/library, ../lib/hipblaslt/library/<arch>). cp -r SRC DST/ with a missing DST creates DST as a copy of SRC and flattens it one level, and the failure is invisible at startup: the engine loads 29 GiB of weights, answers /health, and dies on the first request with rocBLAS error: Cannot read .../lib/rocblas/library/TensileLibrary.dat: No such file or directory for GPU arch : gfx1201. The build script now declares the destination and copies contents.

Model and drafter

File	Bytes	Role
Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf	31,457,990,112	target, 29.29 GiB
Qwen3.8-27B-DFlash2-Q8_0.gguf	2,056,414,816	DFlash2 drafter, 1.92 GiB

Fetched by scripts/download-frank-medium-gguf.sh (resumable, byte-verified — a truncated 31 GB GGUF is otherwise a load-time crash).

The target is Unsloth Dynamic 3.0 (per-tensor types from Unsloth's own quant of the same size, on Unsloth's importance matrix) of an abliterated Swift fine-tune: orcarouter's single refusal direction projected out of 131 tensors, vision tower included. Vision has been ON since 2026-09-17 (--mmproj); this is one of the fleet's two image endpoints.

The file also carries the model's MTP head (blk.64.nextn.*), and since 2026-09-17 that is the deployed drafter: --spec-type draft-mtp builds the draft context against the main model's own weights (shares_model), so there is no second copy of the model and no --model-draft argument — the cost is one nextn layer's KV (4096 B/position) plus small buffers on the card that holds the last layers (ROCm1). The DFlash2 sidecar stays on disk and in the fetch script: SPEC_TYPE=draft-dflash (with the sidecar back on the command line) brings the measured-faster chain back, at the price of ~1.1 GiB on ROCm1.

Deployment

File	Role
scripts/frank-medium.env	every tunable, with the sizing arithmetic in comments
scripts/serve-frank-medium.sh	env-driven llama-server launcher; no arguments
scripts/frank-medium.service	ROOT systemd unit, User=dank, WantedBy=multi-user.target

./scripts/deploy.sh dankserver --service frank-medium

The handler verifies the runtime and payload exist (and refuses otherwise), retires the three vLLM quadlets into ~/.AGENTTRASH/20260916-vllm-radiance-retire/, waits for port 7770 to be released (the retired proxy held it on the host network), relinks and restarts the unit, and gates on http://127.0.0.1:7770/health.

Deployed flags:

Flag	Value	Why
--split-mode / --device / --tensor-split	layer / ROCm0,ROCm1 / 32,33	both cards, one model, no collective. Balanced (operator call 2026-09-19): 65 layers split as evenly as they can be, ROCm1 (the last-layers card that also carries the drafter and output layers) holding the extra one. The predecessor's 35:30 skew (2026-09-17, part 3) leaned work onto the x16 card because ROCm1's Gen4 x1 chipset path made it the 98 % bottleneck; the balanced point trades that lean for symmetry. Split granularity is one LAYER — 0.43 GiB on a DeltaNet boundary, 1.43 on an attention one
--gpu-layers / --fit	999 / off	everything on the cards; auto-fit must not resize a context we set
--ctx-size / --parallel	327680 / 2	the pool over two slots of 163840; --parallel divides the context. 163840 is inside the checkpoint's trained 262144 window, so the whole pool is trained territory — no rope extrapolation (unlike the predecessor's 393216)
--batch-size / --ubatch-size	6144 / 512	measured: ubatch 512 is the prefill optimum and the cheapest point on the curve; the logical batch is throughput- and VRAM-neutral. See frank-medium-batch-ubatch
--ctx-checkpoints / --cache-ram	0 / 16384	no per-slot KV snapshots (the build's default is 32 — this is a deliberate deviation, not an inherited default); the RAM-side prompt cache is capped at 16 GiB, which is what lets a full 163840-token slot state be cached at all (2026-09-18, see below)
--flash-attn / cache types	on / f16 K and V	an agentic checkpoint is the wrong place to spend KV precision
--spec-type / --spec-draft-n-max	draft-dflash / 7	the z-lab DFlash2 Q4_K_M sidecar (1.06 GiB), the measured throughput winner (43.5 tok/s vs the model's own MTP head at 32.4); n_max 7 is the DFlash2 optimum. The target file also carries its own MTP head (nextn_predict_layers=1), so SPEC_TYPE=draft-mtp is one change away — the drafter file stays on disk
--spec-draft-device	ROCm1	the drafter rides the card that ends up holding the target's last layers
--temp / --top-p / --top-k	0.7 / 0.95 / 20	the values the vLLM predecessor served on this endpoint
--reasoning-format / --reasoning-budget	deepseek / -1	trace into reasoning_content, unlimited
--mmproj	mmproj-Qwen3.8-27B-Uncensored-f16.gguf (beside the target)	vision enabled; this build has NO autoload, so the flag is explicit and the serve script refuses to start without the file

The DSpark excursion (2026-09-16, tried and reverted)

The endpoint went live on z-lab DFlash2, was switched to magnitudedev/Qwen3.8-27B-DSpark for a few hours the same day, and was put back. Both are 5-layer block-diffusion sidecars, and their differences are what decided it:

	DFlash2 (deployed)	DSpark (tried, reverted)
file	2,056,414,816 B — 1.92 GiB	1,455,376,576 B — 1.36 GiB
draft depth	block_size 8 -> n_max 7	block_size 7 -> n_max 6 (llama.cpp asserts n_max <= block_size - 1)
attention	sliding_window 2048, causal false, candidate selector (rank 256, top-k 16)	no sliding window; a Markov head in place of the selector
draft KV at 262144	~40 MiB — the window bounds it	5.0 GiB — the whole context

A fetch-script bug this threw up. The DSpark file's size was recorded as 1,359,284,737 when it was fetched — 96 MB short of the Hub's 1,455,376,576 — and that wrong number became the fetch script's declared size, so --verify-only failed on a payload that was in fact byte-correct. Both drafters' sha256 have since been checked against the Hub's LFS oids and match.

The draft context is sized from the target's. cparams.n_ctx = llama_n_ctx(ctx_tgt) in common/speculative.cpp, with no flag to cap it independently, so a drafter that declares no sliding window caches the entire target window on whichever device --spec-draft-device names. DSpark's KV is 5 layers x 2 x 8 kv heads x 128 dim x 2 B = 20 KiB per position.

That cost has to be paid somewhere, and the failure was instructive: at a 1:1 split the model loaded fine, answered /health, and then died on the first draft decode with ROCm error: out of memory inside common_speculative_impl_draft_dflash::process — ROCm1 sat at 33.9 GB of 34.2 GB. A unit that crash-loops with an allocation failure after a successful load is a sizing failure, not a broken drafter.

Read sliding_window out of a drafter's GGUF header before sizing a spec-decode deployment — 40 lines of struct parsing is enough, and the flag-level configuration looks identical either way.

The split moved twice: once for the batch, once for the drafter

Under DSpark. Tripling the batch (2048/512 -> 6144/1536) made the two cards diverge: at 262144 the bigger --ubatch-size cost ROCm1 ~3.3 GiB against ROCm0's ~1.7 GiB — ROCm1 holds the output layers and the drafter's graphs — which left the draft-carrying card at 0.08 GiB free against ROCm0's 3.78 under the 38:27 split that had been balanced at the smaller batch:

-ts	GPU 0 free	GPU 1 free
38,27	3.78 GiB	0.08 GiB
41,24	1.47 GiB	2.38 GiB
40,25	1.88 GiB	1.88 GiB

Then DSpark came off, which frees ~4.3 GiB on ROCm1 (5.0 GiB of draft KV out, 0.65 GiB of heavier DFlash2 weights in) and moves the balance the other way. Re-measured at the same batch, the same two slots and the same window:

-ts	GPU 0 free	GPU 1 free	gap
39,26	2.42 GiB	7.20 GiB	4.78
38,27	3.85 GiB	5.77 GiB	1.92
37,28	4.28 GiB	5.34 GiB	1.06
36,29 — deployed	4.71 GiB	4.91 GiB	0.20
35,30	5.14 GiB	4.48 GiB	0.66
34,31	6.57 GiB	3.05 GiB	3.52

One -ts step is one layer, and this stack's layers are not the same size. qwen35 is 16 x (3 Gated DeltaNet + 1 Gated Attention), so a step is worth ~0.43 GiB on a DeltaNet boundary and ~1.43 GiB on a full-attention one (0.45 GiB of weights plus ~1.0 GiB of its own f16 KV at 262144). Both fell out of that table — 37,28 -> 38,27 moves a DeltaNet layer, 38,27 -> 39,26 an attention layer. Balance by measurement, never by interpolating.

Then the batch moved, and the split followed it (2026-09-16, the same day — see frank-medium-batch-ubatch). 36:29 was the balance at ubatch 1536; the sweep showed 512 is both faster and 3.71 GiB cheaper, and — because the compute buffers cost ROCm1 nearly twice what they cost ROCm0 — shrinking them handed ROCm1 1.31 GiB more than it handed ROCm0:

-ts	ubatch	GPU 0 free	GPU 1 free	gap
36,29	1536	4.71 GiB	4.91 GiB	0.20
36,29	512	6.01 GiB	7.32 GiB	1.31
35,30 — deployed	512	6.44 GiB	6.90 GiB	0.46

35:30's 0.46 GiB gap is one DeltaNet layer's worth, which is the floor here: no split can be tuned finer than a layer, so the 0.20 GiB of the old configuration is not reachable at this batch. It clears the operator's 1.5 GiB-per-card floor by ~5 GiB, so no headroom is bought out of CTX_SIZE and the model's full 262144-token window stays.

A batch of 8192/2048 failed here under DSpark — its 2020 MiB pp buffer could not be allocated on ROCm1, which had 0.08 GiB free. That was a contiguity failure, not a total-bytes one, and the sweep has since removed the reason to care: with the batch geometry measured, there is nothing to gain by making either knob larger, so the question is closed rather than pending.

A fault to watch. A 39:26 split, at 1x batch and one slot, died on its eighth request: the amdgpu driver logged SDMA0 UTCL2 page faults and the engine surfaced ROCm error: an illegal memory access was encountered -> ggml_abort — a fault in the cross-card boundary copy, which is this layer split's hot path. It has not reproduced (36:29 served the whole verification battery and the e2e gate, and 35:30 at ubatch 512 is the deployed shape now; 39:26 itself was only loaded at the deployed batch, not exercised), and it is not an OOM: no amount of headroom would prevent it.

2026-09-17: the pool grows to 327680, the split moves to 36:29, vision comes on

The window grew from 262144 to a 327680-position pool (163840 per conversation over the two slots), the split moved one layer to ROCm0 (35:30 → 36:29), and the BF16 mmproj (888 MiB) came on.

Why 36:29, not the balance. At this context the fit model puts the balance point at 34:31 (0.11 GiB gap). The operator moved a layer to ROCm0 to buy headroom on ROCm1 — the card that carries the drafter, its KV, the output layers and the heavier compute buffers — accepting a 3.15 GiB gap by construction:

Card	Fit	Measured at load	Headroom
ROCm0 (36 layers + mmproj)	29.19 GiB	28.40 GiB	3.46 GiB
ROCm1 (29 layers + drafter)	26.04 GiB	26.66 GiB	5.20 GiB

Both clear the 1.5 GiB floor with room; the fit lands within 0.8 GiB on both cards, which is the accuracy it has always had (at the 35:30 point it reproduced ROCm1 to 0.00 and ROCm0 to 0.86). The seam's slack is the 3.15 GiB gap rounded up, with the trade written in its comment.

The rope caveat. The GGUF carries no context_length and no rope-scaling keys (freq_base 1e7, dimension_sections [11,11,10,0]); the checkpoint's trained window is 262144. Positions beyond that — the top 65k of the pool — extrapolate on the base rope. It runs, and 1.25x is a modest ask of a rope, but it is untrained territory, and the seam's ctx cap encodes 327680 with that note.

Vision: this build has no mmproj autoload. The first 327680/36:29 deploy came up healthy and text-only: image requests answered 500 "image input is not supported — hint: … provide the mmproj" and the load log had no mmproj line, so llama-server fb27a525 does not pick up mmproj*.gguf from the model directory. The serve script now passes --mmproj explicitly (the frank-small pattern), refuses to start without the file when vision is on, and the deploy handler's preflight checks for it. Verified end-to-end after the fix: a generated 64x32 test image was described from its pixels — with thinking on the model reasoned about "the three horizontal bars" (its reading of the crude glyph) and with thinking off answered the letter question in one token.

The e2e crawl spec flaked once on the first 327680 deploy (1 failed / 31 passed) and passed 32/32 on the redeploy; the service itself was healthy throughout (health gate green, NRestarts=0).

2026-09-17 (part 2): the drafter switches to MTP, the pool grows to 393216

The operator asked for 2x196k contexts, which the DFlash2 sidecar's ~1.1 GiB on ROCm1 was eating. The drafter switched to the model's own MTP head (SPEC_TYPE=draft-mtp, n_max 3 — the measured MTP winner from the same-day A/B: 32.4 tok/s at n_max 3 vs 26.7 at n_max 7, against DFlash2's 43.5), the pool grew 327680 → 393216 (196608 per conversation), and the split moved one layer back to 34:31, the balance point at the new pool.

Card	Fit	Measured at load	Headroom
ROCm0 (34 layers + mmproj)	29.77 GiB	28.91 GiB	2.95 GiB
ROCm1 (31 layers + MTP buffers)	28.32 GiB	30.78 GiB	1.08 GiB

The fit predicted 28.32 on ROCm1; the load came in 2.46 GiB hotter than predicted. The MTP charge is one nextn layer's KV (4096 B/position — the draft context is sized from the target, same law as the DFlash2 one) plus draft buffers the fit only estimated. The seam was recalibrated to the measured value: the MTP buffer constant is now 2,901,231,913 B (the 4.20 GiB total MTP charge on card 1), and the floor dropped 1.5 → 1.0 GiB for this shape — ROCm1's 1.08 GiB of headroom is real. 35:30 failed card 0's floor only at the OLD 1.5 GiB bar (it clears 1.0 with 1.27 GiB to spare); 36:29 fails both gates. Under the recalibrated 1.0 GiB floor 35:30 became reachable — and the split went there the same day (part 3).

The rope caveat deepens: the top 131k of a conversation (196608 vs the trained 262144… i.e. 393216/2 = 196608 per slot, so the top 131,072 positions of a FULL conversation) extrapolates on the base rope. 1.5x is a bigger ask than 1.25x; the seam's ctx cap encodes 393216 with that note.

Verified after deploy: adding speculative implementation 'draft-mtp' with shares_model in the load log, no --spec-draft-model flag, --mmproj present; 207 drafts / 108 accepted in a text run (52 % acceptance, mean len ~1.7); vision, tools and reasoning_effort all pass; e2e 32/32; NRestarts=0.

2026-09-17 (part 3): the PCIe topology, and the split leans to 35:30

lspci -vv on dankserver answers why the split now leans toward ROCm0: the two cards do not sit on equal roads.

Card	Path	Bottleneck link
ROCm0 03:00.0	CPU root port 00:01.1 (Gen4 x16) -> switch -> card	Gen4 x16 (~31.5 GB/s)
ROCm1 0b:00.0	CPU root port 00:02.1 (Gen4 x4) -> chipset -> 06:09.0 -> 09:00.0 -> card	Gen4 x1 (~1.97 GB/s)

With the MTP drafter, the output layers and the heavier compute buffers all pinned to the last-layers card, live rocm-smi under traffic read 98% busy on the x1 card vs 40% on the x16 card. The boundary crossing itself is cheap (~5.2 MB of activation per 512-token ubatch rides the wire — a few ms at 1.97 GB/s against a ~430 ms ubatch), but the saturated card should not be the one behind the narrowest link. The operator moved one layer to ROCm0: 34:31 -> 35:30. The moved layer is 34 — the 4k+2 position of block 8, a DeltaNet — which is what keeps the shift cheap (~0.43 GiB).

Fit at the new shape: card 0 30.59 GiB (free 1.27), card 1 30.01 GiB (free 1.85), gap 0.58 — inside the 1.0 GiB floor and the 1.483 GiB slack. The next step, 36:29, crosses the full-attention layer at 35 and fails BOTH gates (card 0 free 0.45 GiB, gap 2.22 GiB), so the layer-toward-ROCm0 lever is spent: further rebalancing means a ctx cut, not a layer. The seam's balance test re-derives all of this from the env file on every run.

Verified after deploy: rocm-smi reads 30.84 GiB used on ROCm0 (headroom 1.02) and 28.86 GiB on ROCm1 (headroom 3.00) — the shift landed 1.80 / 1.97 GiB, more than the fit's 0.58 GiB gap predicted, and the x1 card's headroom tripled. Utilization under live traffic: 60% on the x16 card (was 40%), 98% on the x1 card — the drafter, the output layers and the heavier compute buffers are structurally pinned to the last-layers card, so one DeltaNet layer was all the split lever had left to give. A chat completion answers in 1.4 s end-to-end; MTP acceptance runs 0.51–0.91 (mean len 2.7–3.7) on the live traffic.

2026-09-18: the prompt cache doubles to 16 GiB

--cache-ram is the one cache tunable on this service that is not VRAM: it caps the total bytes of saved prompt states in host RAM (server_prompt_cache, tools/server/server-task.cpp). update() evicts oldest-first past the cap, and alloc() refuses outright any single state larger than the whole cap — with one WRN line and nothing else:

 - prompt state size 11388.990 MiB exceeds cache size limit 8192.000 MiB, skipping

That line appeared 137 times in the two days from 2026-09-16: at 8192 MiB every conversation past ~170k tokens lost its state on the next slot reassignment and had to be re-prefilled in full. It matters more here than the cap alone suggests — this is a hybrid model (16 x 3 DeltaNet + 1 gated attention), so a slot's recurrent cell sits at the end of its own prompt and --ctx-checkpoints 0 leaves no checkpoint to roll back to; llama.cpp then takes the "forcing full prompt re-processing due to lack of cache data" path. The RAM-side cache is the only mechanism that can carry a state across slots, and a refused state is simply gone. The cap is a token budget in disguise, because a state is the serialized KV of a whole sequence plus its constant-size DeltaNet state plus the MTP draft context's own KV:

Component	Per token
16 full-attention layers x 4 kv heads x 256 dim x 2 B x 2 (K,V)	64 KiB
MTP draft context (one nextn layer, same arithmetic)	4096 B
state law	68 KiB

Cap	Tokens storable	Fits one 196608-token slot (~12.8 GiB)?
8192 MiB (until 2026-09-18)	~120k	no — refused, then re-prefilled
16384 MiB (deployed)	~240k	yes, with ~3 GiB of budget spare

The cost is host RAM, allocated on demand and never pinned, so no VRAM figure above moves. dankserver has 62 GiB beside the media-stack co-tenants, which is why the cache is bounded rather than set to -1 (unlimited): the seam test in tests/frank_medium_unit.bats gates both ends — the cap must clear one full slot's state, and it may not claim more than half the host's RAM.

Verified serving behaviour (2026-09-16)

• Load: model loaded ~25 s after start with a warm page cache, then listening on http://0.0.0.0:7770. The drafter adds ~2.5 s.

• Slot shape: n_slots = 2, n_ctx_slot = 196608; /props reports modalities: {vision: true} and the sampler defaults above.

• The drafter is engaged — the init lines are the proof, not the flags. Current (MTP, since 2026-09-17 part 2):

  common_speculative_impl_draft_mtp: adding speculative implementation 'draft-mtp'
  common_speculative_impl_draft_mtp: - n_max=3, n_min=0, p_min=0.00
  ... (draft context shares the main model's weights: `shares_model`)
  

  Under the DFlash2 era (2026-09-16 through 2026-09-17 part 1) the block was:

  common_speculative_init_result: loading draft model '/home/dank/models/Qwen3.8-27B-DFlash2-GGUF/Qwen3.8-27B-DFlash2-Q8_0.gguf'
  common_speculative_impl_draft_dflash: adding speculative implementation 'draft-dflash'
  common_speculative_impl_draft_dflash: - n_max=7, n_min=0, p_min=0.00
  common_speculative_impl_draft_dflash: - block_size=8, mask_token_id=248070, n_extract=5, sample_from_anchor=true
  

  block_size=8 and n_extract=5 are DFlash2's own metadata; the candidate selector is what distinguishes it from DFlash 1. DSpark logs under the same draft_dflash symbol with block_size=7 and a Markov head, so read the numbers, not the implementation name.

• Both cards hold the model. The measurement below is the 34:31-era record of the part 2 cutover; the deployed 35:30 shape is measured in the part 3 section (ROCm0 30.84 GiB / ROCm1 28.86 GiB, headroom 1.02 / 3.00 GiB, both above the 1.0 GiB floor): at 393216 with the deployed batch and the 34:31 split, rocm-smi reported 28.91 GiB used on ROCm0 and 30.78 GiB on ROCm1 of 31.86 GiB each — the fit predicted 29.77 / 28.32, so ROCm1 ran 2.46 GiB hotter than predicted and the seam's MTP constant was recalibrated to the measured value (see the 2026-09-17 part 2 section). ROCm1's 1.08 GiB of headroom cleared the 1.0 GiB floor that shape required.

• Drafting is live on both slots (the 2026-09-16 DFlash2-era numbers below are the record; MTP-era acceptance is in the part 2 section), which is what a 2-slot deployment has to prove: slot print_timing reported draft acceptance = 0.62857 (44 accepted / 70 generated), mean len = 5.40 and 0.48214 (27 / 56) on slot 1, and 0.38776 (57 / 147), mean len = 3.71 on slot 0, from one concurrent pair. Acceptance moves with the text; the point is that neither slot is idle.

• 140,069 prompt tokens prefilled in 134 s (1043 tok/s) was measured through a single slot, at the previous configuration: the raised window is real rather than declared, and it is 22 % beyond the 114688 the vLLM predecessor could offer. The window is now split over two slots.

• Nothing is replicated. The loader does not print a per-device ledger at this verbosity, so this is measured instead of read: raising CTX_SIZE from 131072 to 262144 added 4.83 GB to GPU 0 and 4.84 GB to GPU 1 — both cards carry the KV for their own attention layers, about half each — and neither card's total usage (24.94 / 29.42 GB at 262144) could hold the 31.46 GB file, so no card carries a full copy of the weights. Compare the shape this replaced: two vLLM replicas, i.e. two complete copies of the model, one per card, which is exactly why a 29.8 GB FP8 checkpoint could not be served there at any usable context.

Measured throughput (2026-09-16, deployed service, temperature 0)

Same filler corpus for every shape, so these are shapes rather than a benchmark: a large prompt with a tiny budget isolates prefill, a small prompt with a long budget isolates decode. draft/acc are per request; acceptance moves with the text (filler prose accepts far worse than code or a real agent loop).

Shape	Prefill	Decode	Acceptance
4k prompt / 32 out	3994 tok in 3.25 s — 1230 tok/s	59.8 tok/s	71 %
1k prompt / 512 out	998 tok in 1.41 s — 707 tok/s	43.8 tok/s	41 %
16k prompt / 16 out	12,528 tok in 8.94 s — 1402 tok/s	—	100 % (13/13)
64k prompt / 16 out	48,508 tok in 39.8 s — 1219 tok/s	—	53 %
140k prompt / 16 out	140,069 tok in 134.3 s — 1043 tok/s	—	—
same 1k prompt again	4 tokens prefilled	43.8 tok/s	41 %

The last row is the prefix cache doing its job: the second request of the same conversation re-prefilled 4 of 998 tokens. Prefill holds ~1.0-1.4k tok/s from 4k to 140k, and decode sits at ~44 tok/s on a long generation — against the vLLM predecessor's 42.8 tok/s single-stream (two replicas) and the ROCmFPX llama.cpp measured here at 35.7-46.2 tok/s. One process, one model, both cards, no proxy.

What the drafter is worth — and why MTP is not the alternative

Measured 2026-09-16 with scripts/bench-spec-ab.py on the DFlash2 drafter — the one deployed now: four configurations of the same binary, identical prompts at temperature 0, unit stopped between runs (one host, one copy of a 29 GiB model). 1k-token prompt, 512-token budget. These are the numbers that selected a trained multi-token drafter over the model's own MTP head.

Configuration	Decode	Drafting	VRAM (card 1)
none — autoregressive	17.7 tok/s	—	26.07 GB
draft-mtp n_max 3 (the model's own head)	32.4 tok/s	58 % of 72	29.35 GB
draft-mtp n_max 7	26.7 tok/s	33 % of 147	29.65 GB
draft-dflash n_max 7 — deployed	43.5 tok/s	41 % of 119	29.42 GB

• DFlash2 is worth 2.46× over autoregressive and 34 % over the best MTP. Deeper MTP drafting costs more than it returns: at depth 7 only a third of the proposals survive, at depth 3 more than half do.
• All four produced identical greedy text (2, 3, 5, 7, 11): speculation buys speed, never a different answer. The harness warns loudly if a future configuration ever breaks that.
• MTP is not the cheaper option. Selecting it loads blk.64.nextn.* from the target and adds ~3.3 GB to the card that holds it, against 2.06 GB for the drafter's file — which is also why the load log says unused tensor blk.64.nextn... ignoring under draft-dflash: correct behaviour, not a truncated download.
• They cannot be chained. --spec-type draft-mtp,draft-dflash dies at load with context type MTP requested but model doesn't contain MTP layers. A chain has ONE draft context, and MTP builds against whatever model that context holds — here the DFlash2 file. MTP's head is loaded only when MTP is the draft model (creating MTP draft context against the target model ...). That is structural, not a tuning question.

Chaining the draftless n-gram matcher (deployed)

ngram-mod can ride beside a context-backed drafter, and it is enabled. The reason is in common/speculative.cpp: the n-gram types are added to the config vector unconditionally while every model-backed drafter is gated on ctx_dft != nullptr — so n-gram needs no draft context and cannot collide with the drafter over one. (MTP is the opposite: it is context-backed, which is why draft-mtp,draft-dflash cannot load.) The vector's fixed order — ngram types first, then the drafters — also means CLI order is irrelevant.

The measurements below are DFlash2's — the deployed configuration. (During the DSpark excursion the chain was draft-dspark,ngram-mod; the matcher's behaviour is a property of the matcher, not of the drafter beside it.)

Measured with scripts/bench-spec-ab.py, decode tok/s, three shapes that a proposer can and cannot match:

Configuration	prose 1k/512	echo 1.5k/400 (copy 0.94)	structured/400	identity
draft-dflash@7	43.77	76.82	91.08	70.37
ngram-mod alone	17.72 (0 drafts)	102.50	17.80 (0 drafts)	17.77
draft-dflash,ngram-mod@7	43.65	200.58	91.22	70.40

• On prompt-echo traffic the chain is 2.6× the drafter alone (200.6 vs 76.8) and twice ngram-mod by itself, at 100 % acceptance.
• On prose it is indistinguishable from DFlash2 alone — the same 119 drafts and 49 accepts, the same 43.7 tok/s. That is the property that makes it safe as a default: where the matcher cannot match, it costs nothing.
• ngram-mod alone is a negative control: zero drafts on prose and identity (17.7 tok/s = the autoregressive baseline), 100 % acceptance on echo. Nothing else could have proposed those tokens, which is how its drafts are identified at all — see the trap below.
• VRAM is unchanged (29.42 GB on card 1 with or without it): the matcher is a 64 MiB host-RAM hash table.
• structured (25 records with unique ids) rewards no proposer: its lines differ, so there is no long span to match. It is a negative control for self-repetition, not evidence about it.

The window matters, and shorter is not better. The chain at three windows:

SPEC_NGRAM_N_MIN/N_MAX	prose 1k/512	echo 1.5k/400
48/64 — deployed	43.9	200.6
16/32	36.0	161.0
8/16	36.2	113.8

A short window makes the matcher propose on ordinary prose fragments, where the proposals are rejected and the verification is still paid — the "ngram hurts" result this fleet measured twice before on other payloads. At 48/64 the chain keeps DFlash2's exact numbers on prose and wins on echo.

Operational trap: ngram-mod prints no startup banner in this build. journalctl -u frank-medium shows adding speculative implementation 'draft-dflash' and nothing for the matcher — the only evidence it is running is behavioural (draft_n > 0 where text repeats) or the per-task line slot print_timing: ... draft acceptance = 1.00000 (256 accepted / 256 generated), mean len = 65.00. Do not conclude from the load log that the chain is not engaged.

The agent-loop shape (fleet harness, 2 families × 2 turns)

scripts/bench-turn-latency.py http://127.0.0.1:7770 2 2 LAYERSPLIT — two agent loops re-sending a ~30k-token document concurrently:

	turn 1	turn 2
family 1	24,799 ms	30,697 ms
family 0	51,424 ms	8,015 ms

Read that against the prefill table, not as a latency regression: with one slot the second family's first turn waited for the first (25 s of cold 30k prefill + the wait ≈ 51 s), and the last number — 8.0 s — was a re-prefill the cache could only partly serve. That was the 1-slot configuration. The deployment now runs two slots of 131072, so a second agent loop has a slot of its own instead of queueing, at the cost of halving the window each one gets. --parallel splits the context and the KV between slots; it does not add throughput, so the trade is concurrency against conversation length.

Benchmarking this endpoint

scripts/bench-batch-ubatch.sh sweeps --batch-size against --ubatch-size with llama-bench and records, per cell, the prefill rate, the decode rate and each card's peak VRAM. Run it from a transient unit rather than a shell:

sudo systemd-run --unit=fm-bench --collect --no-block --property=Type=oneshot \
     --property=LimitMEMLOCK=infinity --property=User=dank \
     scripts/bench-batch-ubatch.sh

Properties of llama-bench shape that harness, and three of them have already produced wrong numbers on this fleet:

• It separates devices differently from llama-server. / splits devices within one configuration, , splits sweep values — so this service's --tensor-split 36,29 is -ts 36/29 to llama-bench, and the comma form silently benches each card on its own instead. Likewise --device ROCm0,ROCm1 → -dev ROCm0/ROCm1.
• There is no -c. The context is a consequence of the test shape: n_ctx = n_prompt + n_gen + n_depth. No bench can therefore prove that a configuration's activation buffers fit beside the deployed KV — that needs the live unit at 327680 across two slots.
• A test that cannot allocate calls exit(1), taking the whole process with it, so the harness runs one invocation per cell and a failure costs one TSV line instead of the rest of the sweep.
• A shell's memlock limit is not the unit's. The unit sets LimitMEMLOCK=infinity for the ROCm staging buffers; an SSH shell here gets 8 MiB, which fails the load for a reason that has nothing to do with the model being benched.
• No --spec-* and no --parallel. The drafter cannot take part, so a llama-bench decode figure is the drafterless baseline — 15.9 tok/s at 64k depth against the deployed 43.8 — and slots are never exercised. -d <n> is what puts KV in front of the measurement at all: a tg instance runs with n_prompt = 0, i.e. at ~zero depth unless -d says otherwise.

Observability

The endpoint exports Prometheus metrics under the llamacpp: namespace (--metrics), which is what the fleet's scrape and dashboard read — the vllm: series the predecessor exported no longer exist:

Need	Series
Throughput	llamacpp:prompt_tokens_seconds, llamacpp:predicted_tokens_seconds
Queue	llamacpp:requests_processing, llamacpp:requests_deferred
Cache reuse	llamacpp:prompt_tokens_cached_total vs llamacpp:prompt_tokens_total
Speculative decoding is live	llamacpp:spec_decode_num_drafts_total and llamacpp:spec_decode_num_accepted_tokens_total (their ratio is the acceptance rate)

spec_decode_num_drafts_total staying at 0 is the one-line answer to "is the drafter actually running": the flags can all be right while the draft model failed to load.

Client contract

OpenAI-compatible /v1 on 7770, alias frank-medium, streaming and /health, /props, /metrics as llama.cpp provides them. Reasoning traces are projected into reasoning_content. A request that sends no sampler fields gets the server defaults above; a client that sends its own wins. Images ride the BF16 projector — /props reports modalities: {vision: true, video: true, audio: false} (read live 2026-09-18), and there is no audio anywhere in the fleet.

Troubleshooting

• Loads, then dies on the first request with a rocBLAS Tensile error — the bundled device-library depth (see the bundling trap above); check build-rdna4-rocm714/lib/rocblas/library/ exists and is populated.
• hipErrorNoDevice / a card missing — HIP_VISIBLE_DEVICES must be 0,1; the iGPU enumerates as device 2 and must stay hidden.
• Drafting not engaged — read the journal for common_speculative_impl_draft_dflash (DFlash2 and DSpark both log under that symbol); a missing drafter file makes the serve script exit before llama-server starts, and a drafter whose GGUF the build does not recognise fails the load instead.
• Why is blk.64.* "unused"? That is the model's own MTP head and it is correct for draft-dflash. It is not a truncated download: bash scripts/download-frank-medium-gguf.sh --verify-only proves the bytes.
• A rebuild segfaults at load — the build is gfx1201-only on purpose; a gfx1200 sibling binary links, loads the model and then segfaults on these cards.

See Also

• frank-medium-radiance-mxfp4 — the retired vllm-radiance MXFP4 deployment (replicas + proxy) and the PCIe-atomics evidence
• frank-medium-rocmfpx — the llama.cpp/ROCmFPX deployment earlier on 2026-09-15, whose --split-mode layer measurements this service reuses
• Dual-R9700-Cooperation — why the two cards cannot take part in a collective, and every software escape that is closed
• frank-small-qwen38-gsq-rco — the fleet's other llama.cpp endpoint (RTX 3090, MTP, the image path)
• Dankserver System Specs — hardware inventory
• frank-medium-batch-ubatch — the batch/ubatch sweep that set the deployed geometry and moved this split from 36:29 to 35:30: the ubatch optimum at 512, why the logical batch is free, and the harness traps