| name | deploy-kimi-k26-on-rtx-pro-6000 |
| description | Deploy and serve Moonshot Kimi-K2.6 (1T MoE, MLA, 256K context, vision) in a user-chosen quantization โ official INT4 QAT (moonshotai/Kimi-K2.6, compressed-tensorsโMarlin; vLLM or SGLang) or NVFP4 (nvidia/Kimi-K2.6-NVFP4, ModelOpt FP4; vLLM only โ SGLang NVFP4 is NaN-broken on sm_120) โ on a Linux server (verified Ubuntu 26.04) with 8ร NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB, sm_120) GPUs. The quantization and the engine are both chosen at deploy time with a hardware-based recommendation. Runs an official-image Docker container via nvidia-container-toolkit CDI (--device nvidia.com/gpu=all --ipc=host --network host, bind-mounted weights; --network host is required for NCCL's GPU-to-GPU transport / IB-RoCE GPUDirect RDMA, and the server binds 0.0.0.0 with off-box clients reaching it through an authenticated Caddy proxy that upstreams over loopback 127.0.0.1 โ structural, no firewall), exposing an OpenAI-compatible API on :30000 behind one static systemd service `kimi-k26` (quant + engine selected via its EnvironmentFile โ only one 595 GB variant fits the 8-GPU pool at a time). Use when deploying or serving Kimi-K2.6 INT4 or NVFP4 on RTX PRO 6000 Blackwell / sm_120 hardware (vLLM-in-Docker, or SGLang-in-Docker for INT4) โ or troubleshooting NCCL /dev/shm "unhandled system error" in GPU containers, sm_120 "no kernel image" errors, a missing-`ninja` JIT build failure (-runtime image tag), FlashInfer CuTe-DSL MLIR ICE (llvm.mlir.global_dtors), a vLLM startup ValueError "larger than the available KV cache memory" (gpu-memory-utilization is NOT mem-fraction-static), an OOMโSIGQUIT crash from raising SGLang mem-fraction above 0.85, NVFP4 TRITON_MLA shared-memory OutOfResources at CUDA-graph capture (Required 102400 > limit 101376), NVFP4 "b12x fused MoE requires CUDA 13", NVFP4 offline trust_remote_code FileNotFoundError for a module under blobs/ (e.g. tool_declaration_ts.py), or a slow/hung MoE weight load. |
Deploy Kimi-K2.6 (INT4 QAT or NVFP4) on 8ร RTX PRO 6000 Blackwell Server Edition (sm_120)
Serve Kimi-K2.6 (1T MoE; MLA; 256K; MoonViT vision) in a user-chosen quantization, with an
official-image Docker container โ OpenAI-compatible API on :30000, TP=8, weights bind-mounted
read-only from local NVMe, all in VRAM. Both the quantization and the engine are chosen at
deploy time (steps 2โ3) with a hardware-based recommendation:
- INT4 QAT โ moonshotai/Kimi-K2.6; compressed-tensors โ Marlin (auto); vLLM or SGLang; stock images, no patch. Recommended on sm_120 (official, simplest, both engines verified).
- NVFP4 โ nvidia/Kimi-K2.6-NVFP4; ModelOpt FP4 (
--quantization modelopt_fp4); vLLM only (SGLang NVFP4 = NaN on sm_120). Needs a patched CUDA-13 image (build_nvfp4_image.sh) + offline remote-code prep (prep_remote_code.sh). On sm_120 it gives no throughput win (PCIe-comm-bound: Marlin โ native b12x, native is actually ~12% slower) โ prefer it on datacenter Blackwell (sm_100/B200) where native FP4 (cutedsl) is tuned, or when you specifically need the NVFP4 checkpoint.
Hardware target: 8ร RTX PRO 6000 Blackwell Server Edition (GB202, 96 GB, sm_120) โ ~595 GB of
weights + KV cache need the full 8ร96 GB pool; PCIe-only, no NVLink. Same-chip Workstation/Max-Q
variants should behave identically (unverified). The host needs only the NVIDIA driver (โฅ570, open
kernel module, incl. nvidia-persistenced), Docker, and nvidia-container-toolkit โ no CUDA
toolkit, no Python packages (the download check uses stock python3 + curl).
Why Docker-only (current best practice): the official images ship precompiled sm_120 kernels with
their own CUDA + glibc, so the host-JIT failure class a native venv fights (glibcโฅ2.41 rsqrt
header conflict, ninja, JIT pre-warm against host CUDA) doesn't exist here, and the deploy
reproduces across hosts โ both of those sm_120 fixes are empirically confirmed unnecessary
in-container (image glibc 2.39; the CuTe-DSL norm ICE does not reproduce). GPU access uses
CDI (--device nvidia.com/gpu=all, plain runc โ not the legacy --gpus runtime hook). Treat
step 6 as the go/no-go gate before fronting traffic.
Workflow
-
Verify host โ nvidia-smi: GPU model/count/VRAM (this drives the quantization + engine
recommendations in steps 2โ3); persistence daemon active (systemctl is-active nvidia-persistenced โ ships with the
driver and still matters for containers: it keeps GPU state resident across container restarts;
ad-hoc fallback sudo nvidia-smi -pm 1); ~650 GB free on local NVMe; GPU containers work via
CDI (spec: sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml):
docker run --rm --device nvidia.com/gpu=all --entrypoint nvidia-smi <engine-image>
(Legacy alternative to CDI: --gpus all โ see REFERENCE.md.)
-
Choose the quantization โ ask with AskUserQuestion: "Which Kimi-K2.6 quantization?", options
INT4 QAT and NVFP4 (the tool adds "Other"), marking one (Recommended) by the step-1 GPU:
- sm_120 (RTX PRO 6000 Blackwell โ reference hardware) โ recommend INT4 QAT. Official, both
engines verified, stock images (no patch). NVFP4 here buys no throughput โ the box is
PCIe-comm-bound (no NVLink), so Marlin-of-INT4 โ Marlin-of-NVFP4, and the native FP4 path
(
flashinfer_b12x) measured ~12% slower than Marlin in a 2026-06-11 A/B; it also needs a
patched CUDA-13 image + offline remote-code prep (step 4b). Pick NVFP4 only if you specifically
need that checkpoint.
- Datacenter Blackwell (sm_100 / B200) โ recommend NVFP4 โ native FP4 (FlashInfer cutedsl) is
tuned there; the FP4 tensor cores are the real win. (Not verified by this skill โ upstream path.)
- Hopper / Ada / Ampere (pre-Blackwell, no FP4 tensor cores) โ INT4 QAT (NVFP4 would only dequant).
Sets the checkpoint + flags โ INT4:
moonshotai/Kimi-K2.6 (compressed-tensorsโMarlin, auto-detected,
no --quantization). NVFP4: nvidia/Kimi-K2.6-NVFP4 (--quantization modelopt_fp4, fp8 KV,
--disable-custom-all-reduce, --moe-backend marlin|flashinfer_b12x).
-
Choose the engine โ AskUserQuestion "Which engine?":
- NVFP4 โ vLLM only (don't ask; SGLang NVFP4 = NaN on sm_120, sgl #18954 โ
serve_docker_sglang.sh
hard-refuses it).
- INT4 โ SGLang vs vLLM, mark one (Recommended) by hardware (both verified 5/5, 2026-06-11):
- sm_120 โ recommend SGLang โ faster at every concurrency (+34% @ c1 โฆ +29% @ c128 vs vLLM) and
serves full 256K at mem-fraction 0.85; vLLM here needs util 0.95 and ~131K with bf16 KV.
- Other hardware โ recommend vLLM โ model-card primary path, precompiled kernels, no JIT dep.
The (QUANT, FRAMEWORK) pair is the whole deploy identity; it goes in the env file (step 7), not the
service name.
3b. NVFP4 only โ choose the MoE backend (skip entirely for INT4) โ AskUserQuestion "Which NVFP4
MoE backend?":
- marlin (Recommended on sm_120) โ measured faster at every concurrency (~12% end-to-end;
the box is PCIe-comm-bound, so the native-FP4 GEMM speedup never reaches the wire) and leaves
more KV (b12x reserves extra workspace). The safe throughput pick.
- flashinfer_b12x โ the native FP4 tensor-core path (vLLM PR #40082; needs the CUDA-13
patched image step 4b builds anyway). Pick it to exercise the FP4 tensor cores, or on hardware
where native FP4 is tuned (datacenter Blackwell uses
flashinfer_cutedsl instead). Confirm the
dispatch after launch with scripts/assert-native.sh kimi-k26 (NATIVE FP4 vs MARLIN fallback).
Sets MOE_BACKEND in the env file (step 7); serve_docker_vllm.sh defaults to marlin if unset.
(A/B numbers: the llm-inference-benchmark skill.)
3c. Choose the KV-cache dtype โ AskUserQuestion "Which KV-cache dtype?", options derived from
the (engine, quant) just chosen; mark the first (Recommended):
- SGLang + INT4 โ
fp8_e4m3 (Recommended) โ verified on the reference host: KV pool
2ร (116Kโ232K tokens, full-256K single request fits), verify 5/5 incl. vision, throughput
parity through c64 and +13.5% at c128. | auto (bf16 โ engine default, most conservative
numerics, pool 116K) | fp8_e5m2 (more exponent range, less mantissa โ accepted by SGLang's Triton MLA backend, runtime-unverified here).
- vLLM + INT4 โ
fp8 (Recommended) โ ~2ร KV pool, and required to reach 256K-class
context on 96 GB GPUs (bf16 KV tops out ~131K MAX_MODEL_LEN). | auto (bf16 โ the
verified 0.95/131K config).
- vLLM + NVFP4 โ
fp8 (Recommended โ effectively required): the NVFP4 checkpoint ships
FP8 KV scales; serve_docker_vllm.sh already defaults it for QUANT=nvfp4. | auto (ignores
the shipped scales; bigger KV bytes โ only for debugging numerics).
(Why the vLLM rows list no fp8_e5m2: on sm_120 vLLM serves Kimi through the TRITON_MLA
attention backend, whose KV menu is {auto/bf16, fp8 = fp8_e4m3} only โ fp8_e5m2, nvfp4,
turboquant_* belong to non-MLA backends and hard-fail backend validation at startup. Verified
in-image, vLLM 0.22.1. SGLang uses a different MLA kernel, so its row is governed by its own
supported set โ see the SGLang+INT4 row.)
KV dtype changes numerics โ re-run verify.sh after switching. Sets KV_CACHE_DTYPE in the env
file (step 7). (Measured impact tables: the llm-inference-benchmark skill.)
- Download checkpoint (~595 GB) into the HF hub cache, pinned to a commit. Respect
HF_HOME
(default ~/.cache/huggingface; set HF_HOME to a big-NVMe cache root) โ never hardcode paths. hf/Xet may deadlock โ the script falls back to parallel curl, verifies size/count
vs the paginated HF tree API (curl + python3 stdlib only), writes refs/main:
bash scripts/download.sh moonshotai/Kimi-K2.6 <commit-sha>
bash scripts/download.sh nvidia/Kimi-K2.6-NVFP4 <commit-sha>
4b. NVFP4 only โ build the patched image + fix offline remote-code (skip entirely for INT4):
bash scripts/build_nvfp4_image.sh
bash scripts/prep_remote_code.sh nvidia/Kimi-K2.6-NVFP4
Why: native FP4 MoE needs CUDA 13, and Kimi MLA on sm_120 can only use TRITON_MLA, whose
grouped-decode kernel OOMs at graph capture (smem 102400 > 101376) until the num_stages patch; and
offline trust_remote_code (transformers โฅ5.10) can't resolve the custom module's relative imports
from the symlinked cache (FileNotFoundError: โฆ/blobs/tool_declaration_ts.py). Both fixes are baked
into those two scripts. (REFERENCE.md โ "NVFP4 on sm_120".)
-
Pull the pinned image, then launch (foreground; DETACH=1 = -d --restart unless-stopped).
QUANT selects the checkpoint + flags; the container is always kimi-k26:
docker pull lmsysorg/sglang:v0.5.12.post1-cu130
docker pull vllm/vllm-openai:v0.22.1
QUANT=int4 bash scripts/serve_docker_sglang.sh
QUANT=int4 bash scripts/serve_docker_vllm.sh
QUANT=nvfp4 IMAGE=kimi-k26-nvfp4-vllm:cu130-mla bash scripts/serve_docker_vllm.sh
Container: CDI GPUs, --ipc=host --network host (host-net required for NCCL transport / IB-RoCE
RDMA; server binds 0.0.0.0, Caddy upstreams over loopback โ see REFERENCE "Access model"), weights :ro, memlock/nofile ulimits,
HF_HUB_OFFLINE=1. Load ~10โ15 min from NVMe (~4โ5 min warm cache); ready on "The server is fired up
and ready to roll!" (SGLang) / "Application startup complete" (vLLM) โ docker logs -f kimi-k26.
-
Verify โ health, models, text, tool-call, and vision (sent as a base64 data URL):
bash scripts/verify.sh
bash scripts/assert-native.sh kimi-k26
Vision debug (dumps content + reasoning_content for a red PNG): python3 scripts/vision-probe.py --model kimi-k2.6.
Optional throughput check: use the llm-inference-benchmark skill โ the canonical
bench_sweep.sh, the sweep methodology, and this hardware's recorded baselines all live there.
-
Productionize โ one static kimi-k26.service driven by /etc/kimi-k26.env (selects
FRAMEWORK/QUANT/IMAGE). Only one 595 GB variant fits the 8-GPU pool, so the service name never
changes โ switch quant/engine by editing the env file + sudo systemctl restart kimi-k26 (no
disable/enable). Use the unit or DETACH=1's restart policy, never both. The launcher goes on the
root disk (/usr/local/bin) so the service never depends on a /data mount being present at boot.
sudo install -m755 scripts/serve_docker_vllm.sh scripts/serve_docker_sglang.sh /usr/local/bin/
sudo mkdir -p /var/lib/kimi-k26
sudo cp scripts/kimi-k26.env.example /etc/kimi-k26.env
sudo cp scripts/kimi-k26.service /etc/systemd/system/kimi-k26.service
sudo systemctl daemon-reload && sudo systemctl enable --now kimi-k26
โ Keep /etc/kimi-k26.env values bare โ systemd EnvironmentFile folds an inline # comment
into the value (mangles HF_HOME โ LocalEntryNotFoundError). fp8 KV: add KV_CACHE_DTYPE (vLLM
fp8, SGLang fp8_e4m3) โ verified, ~2ร KV pool.
TLS + Bearer-API-key reverse proxy (engine-agnostic): scripts/setup_proxy.sh. Access model: the
--network host server binds 0.0.0.0, so raw :30000 answers unauthenticated on loopback/
LAN/tailnet โ the auth boundary is Caddy :443 (off-box clients) + the router (no public IP),
not the bind. Caddy upstreams over loopback (127.0.0.1), never the LAN IP โ a LAN-IP upstream
is a DHCP time-bomb (lease moves โ Caddy 502s with an empty body; cost us two outages). No host
firewall (setup_proxy retires any legacy kimi-fw/kimi-netguard). The Caddyfile sets admin off, so apply proxy changes with sudo systemctl restart caddy, not reload. See REFERENCE.md
"Access model".
Key facts (don't relearn these the hard way)
--ipc=host is non-negotiable for TP=8: NCCL needs shared memory and Docker's default 64 MB
/dev/shm breaks it. NCCL "unhandled system error"/SIGBUS right after the load โ check this first.
- INT4 auto-detects (compressed-tensors โ Marlin MoE) โ no
--quantization flag.
- NVFP4 is vLLM-only on sm_120 and needs
--quantization modelopt_fp4 + the patched CUDA-13 image
(build_nvfp4_image.sh): native FP4 (--moe-backend flashinfer_b12x) requires CUDA 13 ("b12x
fused MoE requires CUDA 13") and the TRITON_MLA num_stages smem patch (else graph-capture OOM
102400 > 101376). It genuinely dispatches the FP4 GEMM (no silent dequant) but measured ~12%
slower than Marlin on this PCIe-comm-bound box, so --moe-backend marlin is the throughput pick.
SGLang NVFP4 = NaN on sm_120 โ don't. Offline load also needs prep_remote_code.sh or it dies
with FileNotFoundError โฆ/blobs/tool_declaration_ts.py (transformers โฅ5.10 + symlinked cache).
- Tool calls: SGLang needs only
--tool-call-parser kimi_k2; vLLM needs both
--tool-call-parser kimi_k2 and --enable-auto-tool-choice (missing โ no tool_calls).
- The memory knobs are NOT equivalent across engines (measured): SGLang's
--mem-fraction-static = weights+KV pool with transients outside it โ 0.85 is the verified
ceiling here, 0.90 OOM-crashes the server (vision tower and big-batch transients allocate
outside the pool). vLLM's --gpu-memory-utilization caps the total footprint โ 0.85 leaves
~0.5 GB for KV (won't start); 0.95 is the working setting, and 256K context needs fp8 KV.
FP8 KV (KV_CACHE_DTYPE env in both scripts) doubles the KV pool at zero throughput cost.
- Kimi-K2.6 is a thinking model (reasoning by default) โ answer in
reasoning_content/content;
disable per request with chat_template_kwargs:{"thinking":false}.
- Vision: send images as base64 data URLs. Server-side
image_url URL-fetch gets 403 from
UA-filtering hosts (e.g. Wikimedia) โ fixture issue, not a MoonViT failure. vLLM additionally wants
--mm-encoder-tp-mode data (SGLang needs nothing extra). Debug an empty vision reply with
scripts/vision-probe.py โ it shows whether the answer landed in reasoning_content (a <think>-block
artifact, fixed with more max_tokens) vs a real MoonViT failure.
- MoE weight load is CPU-bound and slow (~10โ15 min); high CPU + 0% GPU + quiet logs = normal loading.
no kernel image is available on sm_120 โ the image predates Blackwell support โ bump the tag.
- Cutover: wait for the GPUs to actually free before relaunching. When swapping variants/engines (or
off an ad-hoc deployment), a ~595 GB / 8-GPU teardown takes 30โ60 s; starting too soon OOMs the workers
at executor init (
Engine core initialization failed, before weight load) and a failed init can
leak GPU memory that turns into a crash-loop. The serve scripts now wait (WAIT_GPU_FREE); if a
loop already leaked, nvidia-smi --query-compute-apps, kill -9 the orphan, confirm GPUs โ 0, restart.
To preserve the client contract across a migration, keep MODEL_NAME stable (it's the OpenAI model id).
- Caddy proxy: loopback upstream +
restart, never reload. setup_proxy.sh points Caddy at
127.0.0.1:30000, not the LAN IP โ the engine binds 0.0.0.0 so loopback works, and it survives
DHCP LAN-IP changes. A LAN-IP upstream is a silent 502-empty-body time-bomb when the host's lease
moves (localhost:30000 keeps working, so it looks like the engine is fine). And the Caddyfile's
admin off disables the :2019 admin API, so systemctl reload caddy fails (dial :2019: connection refused) โ apply Caddyfile/cert/key changes with sudo systemctl restart caddy.
When startup crashes or hangs
See REFERENCE.md: per-engine Docker paths (image pinning, CDI, NCCL/shm, the
SGLang pre-launch gates, the SGLangโvLLM flag map, py-spy load-vs-hang diagnosis),
checkpoint-download gotchas, KV-cache/256K tuning, and productionization (systemd vs restart
policy, proxy, monitoring).