dank/serverAppConfigs
1
0 Fork 0
Code Issues 4 Pull Requests Actions Packages Projects Releases Wiki Activity

large: swap to Kairic Edge IU4 runtime (pinned tag kairic-edge-qwen38-27b-v1)

Browse Source 
Serves jcbtc/Qwen3.8-27B-IU4-Kairic-Edge (four-artifact payload) with
the qualified pinned-runner profile: 262K ctx, 1 slot, b2048/ub512,
t16/tb32, f16 KV, 32 ctx checkpoints, 8 GiB prompt cache with idle-slot
persistence, native MTP4, deterministic sampler, reasoning off.

The serve script replicates the release pinned runner verbatim; the only
differences are safe bind vars (no bare HOST env var; HIP collision rule)
and this machine's split 7.15 SDK LD_LIBRARY_PATH layout.

Rollback: revert this commit and redeploy; v3 model files and the v3
binary are untouched.
This commit is contained in:
Frank committed 2026-08-22 00:21:34 -07:00
1 parent 3b9379bb5b
commit 4c438390dd
2 files changed
+134 -112
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

scripts/frank-default-large.env
+47 -54
View File
@@ -1,40 +1,39 @@
# Tunables for frank-default-large (CIRU llama.cpp, dankclaw, Strix Halo # Tunables for frank-default-large (Kairic Edge llama.cpp, dankclaw,
# gfx1151). # Strix Halo gfx1151).
# #
# Runtime: the PromptForge llama-server built from ciru-ai/ROCmFPX tag # Runtime: the llama-server built from the pinned ciru-ai/ROCmFPX tag
# qwen3.8-activefpx-promptforge-v2.3 plus the v3 output-K8 runtime patch # kairic-edge-qwen38-27b-v1 (Kairic Edge IU4 lane: PromptForge + Dual
# serving jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge (v3). v3 tensor # View + native RDNA 3.5 IU4 matrix compute), built against the private
# allocation (FFN-equalized, late-Q6 writers, full-Q8 output) with the # TheRock ROCm 7.15 SDK kept in the retired ling3-ciru venv.
# full-W8 PromptForge companions, the Output-K8 draft proxy, and the BF16 # Model: jcbtc/Qwen3.8-27B-IU4-Kairic-Edge, four-artifact payload
# vision projector. Alias frank-default-large, port 52625. # (GGUF + FFN/GDN/GDN-Output .pfs sidecars, 26.05 GiB total). Alias
# frank-default-large, port 52625. See docs/kairic-edge.md.
# #
# This runtime replaced the retired vLLM service that previously occupied # This runtime replaced the retired v3 CIRU ActiveFPX PromptForge
# this slot. See docs/ciru-activefpx-promptforge.md. # service. The v3 model files remain in
# /home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3/ and the v3 binary in
# /home/dank/ciru-qwen38/ROCmFPX/build-promptforge/ for rollback.
# #
# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE. The ROCm # WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE. The
# HIP runtime reads those bare env vars (see the 2026-08-13 TEMP incident). # ROCm HIP runtime reads those bare env vars (2026-08-13 TEMP incident).
# llama.cpp reads --host/--port/--temp as CLI flags, never as env vars. # llama.cpp reads --host/--port/--temp as CLI flags, never as env vars.
# CIRU llama-server binary (ROCmFPX v2.3 + v3 output-K8 patch build, see # Kairic Edge llama-server binary (pinned tag build; see
# docs/ciru runtime page). Built against the private TheRock ROCm 7.15 SDK # docs/kairic-edge.md for the build runbook and binary identity).
# kept in the retired ling3-ciru venv. ROCm 7.15 is required for the LLAMA_BIN=/home/dank/ciru-qwen38/ROCmFPX-kairic/build-kairic/bin/llama-server
# advertised Strix Halo performance; the system ROCm 7.2.4 selects
# slower/fallback kernels.
LLAMA_BIN=/home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server
# Model files (v3). The five-file payload (GGUF + FFN/GDN .pfs + Output-K8 # Model files (Kairic Edge four-artifact payload). The sidecars are
# proxy + BF16 vision projector) is the full v3 configuration. The v3 files # accelerator-specific prepacked copies of selected weights; they are
# live in a separate directory so the v2 artifacts stay for rollback. # required, not optional.
MODEL_DIR=/home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3 MODEL_DIR=/home/dank/models/Kairic-Qwen3.8-27B-IU4-Edge
MODEL_FILE=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v3-Q8-FFNEQ-LateQ6.gguf MODEL_FILE=Qwen3.8-27B-IU4-Kairic-Edge.gguf
FFN_PFS=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-FFN.pfs FFN_PFS=Qwen3.8-27B-Kairic-IU4-FFN.pfs
GDN_PFS=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-GDN.pfs GDN_PFS=Qwen3.8-27B-Kairic-IU4-GDN.pfs
OUTPUT_K8_PFS=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-Output-K8.pfs GDN_OUTPUT_PFS=Qwen3.8-27B-Kairic-IU4-GDN-Output.pfs
MMPROJ_FILE=Qwen3.8-27B-mmproj-BF16.mmproj
# ROCm 7.15 SDK paths (inside the retired ling3-ciru venv). The serve script # ROCm 7.15 SDK paths (inside the retired ling3-ciru venv). The serve
# exports ROCM_PATH and LD_LIBRARY_PATH from these so the 7.15 libraries # script exports ROCM_PATH and LD_LIBRARY_PATH from these so the 7.15
# precede any system ROCm 7.2.4. # libraries precede any system ROCm 7.2.4.
ROCM_DEVEL=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_devel ROCM_DEVEL=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_devel
ROCM_LIB=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_libraries ROCM_LIB=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_libraries
ROCM_CORE=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_core ROCM_CORE=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_core
@@ -43,38 +42,32 @@ ALIAS=frank-default-large
BIND_ADDR=0.0.0.0 BIND_ADDR=0.0.0.0
LISTEN_PORT=52625 LISTEN_PORT=52625
# Serving: full native 262144 context (n_ctx_train), 1 slot (user request # Serving: the qualified Kairic profile (pinned runner defaults). Full
# 2026-08-19: parallel slots degraded performance). -c is the TOTAL # native 262144 context, 1 slot, batch 2048 / ubatch 512, 16 target /
# context; with PARALLEL=1 it equals the per-slot context. Concurrent # 32 batch threads, f16 target and draft KV, 32 context checkpoints,
# requests queue on the single slot; the server never preempts a running # 8192 MiB prompt cache with idle-slot persistence. Context checkpoints
# generation. f16 KV, context checkpoints, prompt cache, native MTP # are required for this hybrid recurrent model; do not remove CTX_CHECKPOINTS
# depth-4. The Strix Halo is unified memory; rocm-smi reports the 64 GiB # while the prompt cache is enabled.
# budget. KV at 262144 x 64 KiB/cell is ~16 GiB.
CTX_SIZE=262144 CTX_SIZE=262144
PARALLEL=1 PARALLEL=1
BATCH_SIZE=4096 BATCH_SIZE=2048
UBATCH_SIZE=4096 UBATCH_SIZE=512
TARGET_THREADS=16
BATCH_THREADS=32
CACHE_TYPE_K=f16 CACHE_TYPE_K=f16
CACHE_TYPE_V=f16 CACHE_TYPE_V=f16
CTX_CHECKPOINTS=16 CTX_CHECKPOINTS=32
CACHE_RAM=8192 CACHE_RAM=8192
# Native MTP draft (depth-4, p_min 0, backend sampling). The qualified v3 # Native MTP draft (depth 4, p_min 0, backend sampling) — the pinned
# profile drives the draft via the Output-K8 proxy; the CPU argmax fast # runner's explicit MTP4.
# path and IU4/ngram lanes are disabled in the serve script.
SPEC_DRAFT_N_MAX=4 SPEC_DRAFT_N_MAX=4
SPEC_DRAFT_N_MIN=0 SPEC_DRAFT_N_MIN=0
SPEC_DRAFT_P_MIN=0.0 SPEC_DRAFT_P_MIN=0.0
SPEC_DRAFT_P_SPLIT=0.10 SPEC_DRAFT_P_SPLIT=0.10
# Sampling (model card defaults). # Sampling: the pinned runner keeps the deterministic evaluation
TEMPERATURE=0.7 # profile (temp 0 / top-p 1 / top-k 0 / min-p 0) as the server default.
TOP_P=0.95 # Clients override per request; the model card recommends Qwen live
# values for ordinary chat (temp 0.7, top-p 0.8, top-k 20, min-p 0,
# Thinking enabled (user request, 2026-08-18). deepseek format moves the # presence_penalty 1.5). Reasoning is off in the qualified profile.
# <think> reasoning into message.reasoning_content (OpenAI-compatible);
# none would leave the thought tags unparsed in content.
REASONING=on
REASONING_FORMAT=deepseek
VERBOSITY=4
scripts/serve-frank-large.sh
+87 -58
View File
@@ -1,19 +1,22 @@
#!/usr/bin/env bash #!/usr/bin/env bash
# Serve jcbtc/Qwen3.8-27B-CIRU-ActiveFPX-PromptForge (v3) as # Serve jcbtc/Qwen3.8-27B-IU4-Kairic-Edge as frank-default-large via the
# frank-default-large via the CIRU ROCmFPX llama.cpp runtime on Dankclaw # Kairic Edge CIRU ROCmFPX llama.cpp runtime on Dankclaw (Strix Halo
# (Strix Halo gfx1151). # gfx1151).
# #
# Runtime: the PromptForge llama-server built from ciru-ai/ROCmFPX tag # Runtime: the llama-server built from the pinned ciru-ai/ROCmFPX tag
# qwen3.8-activefpx-promptforge-v2.3 (commit 103af8b41) plus the v3 output-K8 # kairic-edge-qwen38-27b-v1 (commit e97b324) against the private TheRock
# runtime patch (runtime/qwen38-v3-output-k8-runtime.patch, commit af0eac6 on # ROCm 7.15 SDK kept in the retired ling3-ciru venv. The model is the
# Dankclaw) against the private TheRock ROCm 7.15 SDK (kept in the retired # Kairic Edge GGUF plus the three native IU4 sidecars (FFN, GDN, GDN
# ling3-ciru venv). The model is the v3 GGUF with the full-W8 PromptForge # Output); the runtime routes qualified prompt and multi-token
# companion views (FFN/GDN .pfs), the Output-K8 draft proxy, and the BF16 # verification shapes through AMD RDNA 3.5 native IU4 matrix compute.
# vision projector. Native depth-4 MTP speculative decoding. # See docs/kairic-edge.md for the build runbook and binary identity.
# #
# The model card requires ROCm 7.15 (not the system ROCm 7.2.4) for the # This script replicates the release pinned runner
# advertised Strix Halo performance. LD_LIBRARY_PATH therefore points at the # (ROCmFPX-kairic/scripts/run-kairic-edge-gfx1151.sh) verbatim. The only
# 7.15 SDK libraries first. An older system ROCm must not precede them. # deliberate differences: --host/--port come from BIND_ADDR/LISTEN_PORT
# (safe names; the ROCm HIP runtime reads a bare HOST env var, see the
# 2026-08-13 TEMP incident) and LD_LIBRARY_PATH uses this machine's split
# 7.15 SDK layout.
# #
# Tunables come from frank-default-large.env in this directory unless # Tunables come from frank-default-large.env in this directory unless
# FRANK_LARGE_ENV_FILE overrides that path. # FRANK_LARGE_ENV_FILE overrides that path.
@@ -29,55 +32,40 @@ if [[ -f "$ENV_FILE" ]]; then
fi fi
# --- defaults (used when the env file does not set them) ------------------- # --- defaults (used when the env file does not set them) -------------------
: "${LLAMA_BIN:=/home/dank/ciru-qwen38/ROCmFPX/build-promptforge/bin/llama-server}" # The defaults are the release pinned runner's qualified values.
: "${MODEL_DIR:=/home/dank/models/CIRU-Qwen3.8-27B-ActiveFPX-v3}" : "${LLAMA_BIN:=/home/dank/ciru-qwen38/ROCmFPX-kairic/build-kairic/bin/llama-server}"
: "${MODEL_FILE:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-v3-Q8-FFNEQ-LateQ6.gguf}" : "${MODEL_DIR:=/home/dank/models/Kairic-Qwen3.8-27B-IU4-Edge}"
: "${FFN_PFS:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-FFN.pfs}" : "${MODEL_FILE:=Qwen3.8-27B-IU4-Kairic-Edge.gguf}"
: "${GDN_PFS:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-GDN.pfs}" : "${FFN_PFS:=Qwen3.8-27B-Kairic-IU4-FFN.pfs}"
: "${OUTPUT_K8_PFS:=Qwen3.8-27B-CIRU-ActiveFPX-PromptForge-Output-K8.pfs}" : "${GDN_PFS:=Qwen3.8-27B-Kairic-IU4-GDN.pfs}"
: "${MMPROJ_FILE:=Qwen3.8-27B-mmproj-BF16.mmproj}" : "${GDN_OUTPUT_PFS:=Qwen3.8-27B-Kairic-IU4-GDN-Output.pfs}"
: "${ROCM_DEVEL:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_devel}" : "${ROCM_DEVEL:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_devel}"
: "${ROCM_LIB:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_libraries}" : "${ROCM_LIB:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_libraries}"
: "${ROCM_CORE:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_core}" : "${ROCM_CORE:=/home/dank/ling3-ciru/Ling-3.0-Flash-CIRU-int4-Strix-native/.venv/lib/python3.12/site-packages/_rocm_sdk_core}"
: "${ALIAS:=frank-default-large}" : "${ALIAS:=frank-default-large}"
: "${BIND_ADDR:=0.0.0.0}" : "${BIND_ADDR:=0.0.0.0}"
: "${LISTEN_PORT:=52625}" : "${LISTEN_PORT:=52625}"
: "${CTX_SIZE:=524288}" : "${CTX_SIZE:=262144}"
: "${PARALLEL:=2}" : "${PARALLEL:=1}"
: "${BATCH_SIZE:=4096}" : "${BATCH_SIZE:=2048}"
: "${UBATCH_SIZE:=4096}" : "${UBATCH_SIZE:=512}"
: "${TARGET_THREADS:=16}"
: "${BATCH_THREADS:=32}"
: "${CACHE_TYPE_K:=f16}" : "${CACHE_TYPE_K:=f16}"
: "${CACHE_TYPE_V:=f16}" : "${CACHE_TYPE_V:=f16}"
: "${CTX_CHECKPOINTS:=16}" : "${CTX_CHECKPOINTS:=32}"
: "${CACHE_RAM:=8192}" : "${CACHE_RAM:=8192}"
: "${SPEC_DRAFT_N_MAX:=4}" : "${SPEC_DRAFT_N_MAX:=4}"
: "${SPEC_DRAFT_N_MIN:=0}" : "${SPEC_DRAFT_N_MIN:=0}"
: "${SPEC_DRAFT_P_MIN:=0.0}" : "${SPEC_DRAFT_P_MIN:=0.0}"
: "${SPEC_DRAFT_P_SPLIT:=0.10}" : "${SPEC_DRAFT_P_SPLIT:=0.10}"
: "${TEMPERATURE:=0.7}"
: "${TOP_P:=0.95}"
: "${REASONING:=on}"
: "${REASONING_FORMAT:=deepseek}"
: "${VERBOSITY:=4}"
MODEL="$MODEL_DIR/$MODEL_FILE" MODEL="$MODEL_DIR/$MODEL_FILE"
export PROMPTFORGE_SIDECAR="$MODEL_DIR/$FFN_PFS" FFN="$MODEL_DIR/$FFN_PFS"
export PROMPTFORGE_GDN_SIDECAR="$MODEL_DIR/$GDN_PFS" GDN="$MODEL_DIR/$GDN_PFS"
# v3 qualified runtime: Output-K8 draft proxy + small-row W8 prompt route. GDN_OUTPUT="$MODEL_DIR/$GDN_OUTPUT_PFS"
export PROMPTFORGE_MODE="m2048_fused_tail1476"
export GGML_CUDA_GRAPH_OPT=0
export PROMPTFORGE_ENABLE_SMALLM_W8=1
export PROMPTFORGE_MTP_OUTPUT_K8=1
export PROMPTFORGE_MTP_OUTPUT_K8_VALIDATE=0
export PROMPTFORGE_MTP_OUTPUT_K8_PROXY="$MODEL_DIR/$OUTPUT_K8_PFS"
# The qualified v3 profile disables the CPU argmax draft fast path; the
# Output-K8 proxy + backend sampling replace it. Also drop any IU4/ngram
# companions from the non-IU4 development lane.
unset LLAMA_MTP_CPU_ARGMAX_FASTPATH LLAMA_MTP_BACKEND_GREEDY || true
unset PROMPTFORGE_IU4_SIDECAR PROMPTFORGE_GDN_IU4_SIDECAR || true
unset PROMPTFORGE_ENABLE_NGRAM_MOD PROMPTFORGE_ENABLE_NGRAM_M65_IU4 || true
for f in "$MODEL" "$PROMPTFORGE_SIDECAR" "$PROMPTFORGE_GDN_SIDECAR" "$PROMPTFORGE_MTP_OUTPUT_K8_PROXY" "$MODEL_DIR/$MMPROJ_FILE"; do for f in "$MODEL" "$FFN" "$GDN" "$GDN_OUTPUT"; do
if [[ ! -f "$f" ]]; then if [[ ! -f "$f" ]]; then
echo "missing required file: $f" >&2 echo "missing required file: $f" >&2
exit 1 exit 1
@@ -89,15 +77,52 @@ if [[ ! -x "$LLAMA_BIN" ]]; then
fi fi
# ROCm 7.15 SDK libraries must precede any system ROCm in LD_LIBRARY_PATH. # ROCm 7.15 SDK libraries must precede any system ROCm in LD_LIBRARY_PATH.
# The release runner assumes one ROCm tree; this machine keeps the 7.15
# SDK in three split directories, so use the proven split layout plus the
# server's own directory.
export ROCM_PATH="$ROCM_DEVEL" export ROCM_PATH="$ROCM_DEVEL"
export LD_LIBRARY_PATH="${ROCM_LIB}/lib:${ROCM_CORE}/lib:${ROCM_DEVEL}/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" export LD_LIBRARY_PATH="$(dirname "$LLAMA_BIN"):${ROCM_LIB}/lib:${ROCM_CORE}/lib:${ROCM_DEVEL}/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
# Pinned runner environment, verbatim (Kairic Edge IU4 lane, PromptForge,
# Dual View sidecars).
export HIP_VISIBLE_DEVICES=0
export HSA_OVERRIDE_GFX_VERSION=11.5.1
export GGML_CUDA_GRAPH_OPT=0
export LLAMA_TARGET_GREEDY_ARGMAX_FASTPATH=1
export LLAMA_MTP_CPU_ARGMAX_FASTPATH=1
export PROMPTFORGE_TARGET_ONLY=0
export PROMPTFORGE_MODE=iu4_ffn
export PROMPTFORGE_IU4_SIDECAR="$FFN"
export PROMPTFORGE_SIDECAR="$FFN"
export PROMPTFORGE_IU4_HADAMARD=1
export PROMPTFORGE_IU4_SEGMENTED=0
export PROMPTFORGE_ENABLE_FFN_KEEPERS=0
export PROMPTFORGE_FFN_KEEPERS=late6
export PROMPTFORGE_ENABLE_GDN=1
export PROMPTFORGE_ENABLE_GDN_W8=0
export PROMPTFORGE_GDN_SIDECAR_OVERRIDE="$GDN"
export PROMPTFORGE_GDN_SIDECAR="$GDN"
export PROMPTFORGE_GDN_IU4_HADAMARD=1
export PROMPTFORGE_ENABLE_GDN_KEEPERS=0
export PROMPTFORGE_ENABLE_GDN_OUTPUT=1
export PROMPTFORGE_GDN_OUTPUT_SIDECAR="$GDN_OUTPUT"
export PROMPTFORGE_GDN_OUTPUT_KEEPERS=v3_lateq6
export PROMPTFORGE_ENABLE_SMALLM_IU4=1
export PROMPTFORGE_ENABLE_SMALLM_GDN_IU4=1
export PROMPTFORGE_ENABLE_SMALLM_GDN_OUTPUT_IU4=0
export PROMPTFORGE_ENABLE_IU4_DECODE_M2_M5=0
export PROMPTFORGE_OUTPUT_K8_STRICT_GREEDY=0
# Pinned runner flags, verbatim (qualified 262K one-slot profile, native
# MTP depth 4, deterministic sampler, reasoning off). Sampling is
# overridden per request by clients; the card recommends Qwen live values
# (temp 0.7, top-p 0.8, top-k 20, min-p 0, presence_penalty 1.5).
exec "$LLAMA_BIN" \ exec "$LLAMA_BIN" \
--model "$MODEL" \ -m "$MODEL" \
--mmproj "$MODEL_DIR/$MMPROJ_FILE" \
--alias "$ALIAS" \ --alias "$ALIAS" \
--host "$BIND_ADDR" \ --host "$BIND_ADDR" \
--port "$LISTEN_PORT" \ --port "$LISTEN_PORT" \
--jinja \
-dev ROCm0 \ -dev ROCm0 \
-ngl 999 \ -ngl 999 \
-c "$CTX_SIZE" \ -c "$CTX_SIZE" \
@@ -106,25 +131,29 @@ exec "$LLAMA_BIN" \
-fa on \ -fa on \
-ctk "$CACHE_TYPE_K" \ -ctk "$CACHE_TYPE_K" \
-ctv "$CACHE_TYPE_V" \ -ctv "$CACHE_TYPE_V" \
-t "$TARGET_THREADS" \
-tb "$BATCH_THREADS" \
-np "$PARALLEL" \ -np "$PARALLEL" \
-ctxcp "$CTX_CHECKPOINTS" \ -ctxcp "$CTX_CHECKPOINTS" \
--cache-ram "$CACHE_RAM" \ --cache-ram "$CACHE_RAM" \
--cache-prompt \ --cache-prompt \
--cache-idle-slots \
--metrics \ --metrics \
--no-mmap \ --kairic-edge \
--spec-type draft-mtp \ --spec-type draft-mtp \
--spec-draft-device ROCm0 \ --spec-draft-device ROCm0 \
--spec-draft-ngl 999 \ --spec-draft-ngl 999 \
--spec-draft-type-k f16 \ --spec-draft-type-k "$CACHE_TYPE_K" \
--spec-draft-type-v f16 \ --spec-draft-type-v "$CACHE_TYPE_V" \
--spec-draft-n-max "$SPEC_DRAFT_N_MAX" \ --spec-draft-n-max "$SPEC_DRAFT_N_MAX" \
--spec-draft-n-min "$SPEC_DRAFT_N_MIN" \ --spec-draft-n-min "$SPEC_DRAFT_N_MIN" \
--spec-draft-p-min "$SPEC_DRAFT_P_MIN" \ --spec-draft-p-min "$SPEC_DRAFT_P_MIN" \
--spec-draft-p-split "$SPEC_DRAFT_P_SPLIT" \ --spec-draft-p-split "$SPEC_DRAFT_P_SPLIT" \
--spec-draft-backend-sampling \ --spec-draft-backend-sampling \
--temp "$TEMPERATURE" \ --temp 0 \
--top-p "$TOP_P" \ --top-p 1.0 \
--reasoning "$REASONING" \ --top-k 0 \
--reasoning-format "$REASONING_FORMAT" \ --min-p 0.0 \
--reasoning-budget -1 \ --reasoning off \
-lv "$VERBOSITY" --reasoning-format none \
--reasoning-budget -1