Frank
b6d4914fcc
DSpark declares no sliding window, so its draft context -- built with n_ctx = llama_n_ctx(ctx_tgt) -- caches the whole target window: 5 layers x 2 x 8 kv heads x 128 dim x 2 B = 20 KiB/position, i.e. 5.0 GiB at 262144, where DFlash2's sliding_window 2048 cost ~40 MiB. At a 1:1 split that put ROCm1 at 33.9 GB of 34.2 GB and the first draft graph capture died inside common_speculative_impl_draft_dflash::process with "ROCm error: out of memory" -- after the model had loaded and /health had answered, which is why the deploy looked like a crash-loop rather than a sizing failure. Handing three target layers to ROCm0 restores headroom on both cards at the full window. Measured at 262144: 28.52 GB / 30.71 GB, i.e. 5.30 GiB and 3.26 GiB free (was 2.7 GB free on the draft-carrying card with 1:1 + DFlash2). The unit seam now models the busier card as a skewed share plus the drafter's full-context KV, calibrated to the rocm-smi figure. AGENTS.md gains the durable lesson: read sliding_window out of the drafter's GGUF header before sizing a spec-decode deployment.
225 lines
11 KiB
Bash
225 lines
11 KiB
Bash
# Tunables for frank-medium (ajgazin/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP
|
|
# UD-Q8_K_XL + magnitudedev/Qwen3.8-27B-DSpark-GGUF Q8_0 drafter), mainline llama.cpp
|
|
# on Dankserver, 2x AMD Radeon AI PRO R9700 (gfx1201). Since 2026-09-16 this
|
|
# is the SOLE LLM runtime on the box: the vllm-radiance replica pair and its
|
|
# routing proxy are retired.
|
|
#
|
|
# Hardware: two R9700 (Navi 48, RDNA4, 32,624 MiB each), no XGMI, different
|
|
# IOMMU groups -- every cross-card byte goes over PCIe. The Raphael iGPU
|
|
# (gfx1036/Radeon 610M) enumerates as HIP device 2 and is hidden below so
|
|
# nothing can drift onto shared system RAM.
|
|
#
|
|
# WHY LAYER SPLIT AND NOT TENSOR (operator call, 2026-09-16): tensor parallel
|
|
# needs a cross-card allreduce per layer, and this host cannot dispatch one.
|
|
# RCCL's kernels carry a hidden hostcall-buffer parameter, and ROCr refuses it
|
|
# on a device without PCIe atomics -- the second card sits behind the B650
|
|
# chipset bridges (AtomicOpsCap: Routing-, read-only), so every collective that
|
|
# touches it dies with hipErrorIllegalState. Measured on this box 2026-09-15
|
|
# (llama-bench, pp4096 / tg128, f16 KV, fa on):
|
|
#
|
|
# single R9700 1160.9 / 19.42
|
|
# layer 2x 1991.2 / 18.76 (859.1 / 17.64 at depth 32768)
|
|
# tensor 2x 271.8 / 26.26 (no RCCL: butterfly fallback, -7x prefill)
|
|
#
|
|
# Layer split distributes layers in order; only the activation at the boundary
|
|
# crosses the wire, once per forward pass. That is the "pipeline parallelism"
|
|
# this service wanted, and it needs no collective at all.
|
|
#
|
|
# Runtime: mainline llama.cpp master fb27a525 (2026-09-16), gfx1201, built by
|
|
# scripts/build-llama-rdna4.sh against a private ROCm 7.14.0a20260624 toolchain
|
|
# that is bundled next to the binary (RPATH $ORIGIN/../lib, no LD_LIBRARY_PATH).
|
|
# The ref is pinned by TWO features: DFlash2 (ggml-org/llama.cpp#27342 -- the
|
|
# drafter is detected from its GGUF) and qwen35 MTP. Do not float it.
|
|
#
|
|
# Model: ajgazin/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF @ UD-Q8_K_XL,
|
|
# 29.29 GiB, arch qwen35, Unsloth Dynamic 3.0 (per-tensor types from Unsloth's
|
|
# own quant of the same size, on Unsloth's importance matrix). Abliteration is
|
|
# orcarouter's single refusal direction projected out of 131 tensors; the vision
|
|
# tower is untouched but unused here (text only, see NO_MMPROJ). The file also
|
|
# carries the model's MTP head (blk.64.nextn.*), so a file-level switch to
|
|
# --spec-type draft-mtp is available without a second copy in VRAM.
|
|
#
|
|
# Drafter: magnitudedev/Qwen3.8-27B-DSpark-GGUF @ Q8_0, 1.27 GiB: 5 layers,
|
|
# hidden 5120, block-diffusion proposal with the Markov head (dspark).
|
|
# Decoding stays lossless -- the target verifies every drafted token. Its KV
|
|
# spans the whole context (see SIZING): unlike DFlash2 it declares no sliding
|
|
# window, and that is what the tensor split below pays for.
|
|
#
|
|
# Architecture (GGUF arch qwen35): 64 layers as 16 x (3 x (Gated DeltaNet ->
|
|
# FFN) + 1 x (Gated Attention -> FFN)) -- 48 linear-attention layers whose
|
|
# per-sequence state is CONSTANT-SIZE, 16 full-attention layers at 4 KV heads x
|
|
# 256 head_dim. A token therefore costs 2 x 4 x 256 x 2 B = 64 KiB of KV at f16
|
|
# in the attention layers only, plus ~0.16 GiB per sequence of f32 DeltaNet
|
|
# state at any depth.
|
|
#
|
|
# SIZING, measured (rocm-smi, 2026-09-16). Weights split across the two cards;
|
|
# the drafter AND its KV ride the second. The loader's compute buffers are part
|
|
# of what is seen below -- they are allocated at load, not per request:
|
|
#
|
|
# CTX_SIZE split GPU 0 GPU 1 headroom on the busier card
|
|
# 131072 1,1 19.89 GB 24.30 GB 9.9 GB / 31 % (DFlash2)
|
|
# 262144 1,1 24.72 GB 29.14 GB 5.07 GB / 15 % (DFlash2) <- before
|
|
# 262144 38,27 28.52 GB 30.71 GB 3.50 GB / 10 % (DSpark) <- deployed
|
|
#
|
|
# 262144 is the window the model ships (n_ctx_train) and it fits with room, so
|
|
# the fleet gets the whole thing; the arithmetic that used to cap this endpoint
|
|
# at 114688 was the vLLM predecessor's, not this runtime's. If a future change
|
|
# adds VRAM (a longer drafter, quantized KV, more slots) re-read the numbers
|
|
# rather than assuming, and prefer CTX_SIZE over --gpu-layers.
|
|
#
|
|
# WHY THE SPLIT IS 38:27. The draft context is built with `n_ctx =
|
|
# llama_n_ctx(ctx_tgt)` (common/speculative.cpp), so the drafter's KV scales
|
|
# with the target window, and DSpark -- unlike DFlash2, which declared
|
|
# sliding_window 2048 and cost ~40 MiB -- caches all of it: 5 layers x 2 x 8 kv
|
|
# heads x 128 dim x 2 B = 20 KiB per position, i.e. 5.0 GiB at 262144. At a
|
|
# 1:1 split that put ROCm1 at 33.9 GB of 34.2 GB and the draft's first graph
|
|
# capture died with `ROCm error: out of memory` (in
|
|
# common_speculative_impl_draft_dflash::process). Handing three target layers
|
|
# to ROCm0 buys back ~3 GiB of headroom on both cards at the full window.
|
|
#
|
|
# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE (HIP reads
|
|
# those bare vars; 2026-08-13 incident). llama.cpp takes --host/--port/--temp
|
|
# as flags only.
|
|
|
|
# llama-server: mainline fb27a525 built for gfx1201 by scripts/build-llama-rdna4.sh.
|
|
LLAMA_BIN=/home/dank/GitPackages/llama.cpp-dflash2/build-rdna4-rocm714/bin/llama-server
|
|
|
|
# Payload, fetched with scripts/download-frank-medium-gguf.sh.
|
|
MODEL_DIR=/home/dank/models/Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-GGUF
|
|
MODEL_FILE=Swift-Qwen3.8-27B-Uncensored-Dynamic-MTP-UD-Q8_K_XL.gguf
|
|
DRAFT_MODEL_DIR=/home/dank/models/Qwen3.8-27B-DSpark-GGUF
|
|
DRAFT_MODEL_FILE=Qwen3.8-27B-DSpark-Q8_0.gguf
|
|
|
|
# Fleet identity. 7770 is the fleet's dankserver endpoint; the alias is the id
|
|
# at /v1/models and what every client asks for.
|
|
ALIAS=frank-medium
|
|
BIND_ADDR=0.0.0.0
|
|
LISTEN_PORT=7770
|
|
# Device placement. HIP_VISIBLE_DEVICES hides the Raphael iGPU (HIP device 2)
|
|
# so nothing can drift onto shared system RAM; DEVICES is what llama.cpp splits
|
|
# across, and neither of the two cards has an XGMI link to the other.
|
|
HIP_VISIBLE_DEVICES=0,1
|
|
DEVICES=ROCm0,ROCm1
|
|
SPLIT_MODE=layer
|
|
|
|
# Skewed 38:27 rather than 1:1: ROCm1 also carries the drafter and the whole of
|
|
# its KV, so it must hold fewer target layers for the cards to end level. See
|
|
# SIZING above -- this is measured, not computed.
|
|
TENSOR_SPLIT=38,27
|
|
|
|
NGL=999
|
|
# Auto-fit has no business resizing a context we set explicitly.
|
|
FIT=off
|
|
FLASH_ATTN=on
|
|
|
|
# One slot at the model's full configured window (n_ctx_train 262144), which was
|
|
# measured to fit before it was deployed: see SIZING above. --parallel DIVIDES
|
|
# --ctx-size between slots, so one slot carries the whole window and a second
|
|
# concurrent request queues behind it.
|
|
CTX_SIZE=262144
|
|
PARALLEL=1
|
|
|
|
BATCH_SIZE=2048
|
|
UBATCH_SIZE=512
|
|
|
|
# Threads cover tokenization, sampling and HTTP only: every layer is on a GPU.
|
|
NUM_THREADS=8
|
|
THREAD_BATCH=8
|
|
|
|
# f16 K/V: an agentic checkpoint is the wrong place to spend KV precision.
|
|
CACHE_TYPE_K=f16
|
|
CACHE_TYPE_V=f16
|
|
CTX_CHECKPOINTS=8
|
|
CACHE_RAM=8192
|
|
|
|
# Sampling defaults for requests that send none -- the values the vLLM
|
|
# predecessor served on this endpoint (`--override-generation-config`), so a
|
|
# client sees no change across the runtime swap. The model card's own
|
|
# recommendation is temp 1.0 / top_p 0.95; the harness sets its own per model in
|
|
# models.yml, so this is the floor, not the contract.
|
|
TEMPERATURE=0.7
|
|
TOP_P=0.95
|
|
TOP_K=20
|
|
MIN_P=0.0
|
|
PRESENCE_PENALTY=0.0
|
|
|
|
# Reasoning on, projected into reasoning_content (deepseek format) with no
|
|
# budget: the model decides how long to think.
|
|
REASONING=on
|
|
REASONING_FORMAT=deepseek
|
|
REASONING_BUDGET=-1
|
|
|
|
# Chat template: empty = the one embedded in the GGUF (Qwen3.8's own).
|
|
CHAT_TEMPLATE_FILE=
|
|
CHAT_TEMPLATE_KWARGS='{"enable_thinking":true,"preserve_thinking":true}'
|
|
|
|
# Speculative decoding: the DSpark block-diffusion drafter PLUS the draftless
|
|
# n-gram matcher. DSpark's GGUF carries block_size 7, so at most 6 draft
|
|
# tokens per verification step (llama.cpp asserts n_max <= block_size - 1).
|
|
# The drafter rides ROCm1, the card that ends up holding the target's last
|
|
# layers.
|
|
SPEC_TYPE=draft-dspark,ngram-mod
|
|
SPEC_DRAFT_N_MAX=6
|
|
SPEC_DRAFT_NGL=999
|
|
SPEC_DRAFT_DEVICE=ROCm1
|
|
SPEC_DRAFT_TYPE_K=f16
|
|
SPEC_DRAFT_TYPE_V=f16
|
|
|
|
# MTP (the model's own head) vs the drafter -- MEASURED 2026-09-16 (temperature
|
|
# 0, 1k prompt / 512 out; the autoregressive baseline and both MTP depths were
|
|
# run against the deployed configuration in one session, same binary, same
|
|
# prompts):
|
|
#
|
|
# none 17.7 tok/s card1 26.07 GB
|
|
# draft-mtp n=3 32.4 tok/s 58 % accepted card1 29.35 GB
|
|
# draft-mtp n=7 26.7 tok/s 33 % accepted card1 29.65 GB
|
|
# draft-dflash 7 43.5 tok/s 41 % accepted card1 29.42 GB
|
|
#
|
|
# All four produced IDENTICAL greedy text, so verification is lossless whichever
|
|
# proposer runs. DFlash2 is worth 2.46x over autoregressive and 34 % over the
|
|
# best MTP; deeper MTP drafting costs more than it returns (n=7 accepts 33 % of
|
|
# 147 proposals, n=3 accepts 58 % of 72). The MTP head is not free either:
|
|
# selecting it loads blk.64.nextn.* and adds ~3.3 GB to the card that holds it,
|
|
# against 2.06 GB for the drafter's file. Do not switch this without
|
|
# re-measuring.
|
|
#
|
|
# THE TWO CANNOT BE CHAINED. `--spec-type draft-mtp,draft-dflash` dies at load:
|
|
# context type MTP requested but model doesn't contain MTP layers
|
|
# failed to create MTP context
|
|
# A chain has ONE draft context, and MTP builds against whatever model that
|
|
# context holds -- here the DFlash2 file, which has no MTP layers. MTP's head is
|
|
# only loaded when MTP IS the draft model ("creating MTP draft context against
|
|
# the target model ..."), so a chain of the two is a structural dead end, not a
|
|
# tuning question. See serverAppConfigs-docs/LLM/frank-medium-llamacpp-dflash2.md.
|
|
|
|
# ngram-mod: the draftless n-gram matcher (CPU hash table in host RAM, no VRAM,
|
|
# no model file, no draft context). It joins the chain unconditionally --
|
|
# unlike MTP, which needs a draft context and therefore cannot share one with
|
|
# DFlash2. Verified on the loaded server: with ngram-mod as the ONLY spec type,
|
|
# a prose prompt produces ZERO drafts (plain autoregressive) while a
|
|
# prompt-echo prompt accepts 100 % of them, so its drafts are unmistakable.
|
|
#
|
|
# WINDOW MEASURED 2026-09-16 (chain at each window, decode tok/s):
|
|
#
|
|
# n_min/n_max prose 1k/512 prompt-echo 1.5k/400
|
|
# 48/64 43.9 200.6 <- deployed
|
|
# 16/32 36.0 161.0
|
|
# 8/16 36.2 113.8
|
|
#
|
|
# The long window is the right one, and not only because it wins on echo: a
|
|
# SHORT window makes the matcher propose on short fragments of ordinary prose,
|
|
# where the proposals are rejected and the verification still costs -- which is
|
|
# the "ngram hurts" result the fleet measured on other payloads twice before.
|
|
# With 48/64 the chain is byte-identical to DFlash2 alone on prose (same 119
|
|
# drafts / 49 accepted) and on the identity prompt (70/53).
|
|
#
|
|
# Do not shorten these without re-running scripts/bench-spec-ab.py: the
|
|
# no-penalty property is what makes the chain safe as a default.
|
|
SPEC_NGRAM_N_MATCH=24
|
|
SPEC_NGRAM_N_MIN=48
|
|
SPEC_NGRAM_N_MAX=64
|
|
|
|
# Text only: the server must not autoload a vision tower. Vision stays on
|
|
# frank-small (Dankclaw, RTX 3090), which is the fleet's image endpoint.
|
|
NO_MMPROJ=1
|