GitHub - ryanzhou/deepseek-v4-flash-mi300x

DeepSeek V4 Flash on a single AMD MI300X This repository contains the configuration and patches I use to run deepseek-ai/DeepSeek-V4-Flash-0731 on one AMD MI300X in production. It includes the Docker Compose stack, SHA-256-pinned file overlays, reference diffs against upstream, and tuning tables. The checkpoint runs as shipped, without additional weight quantization or offload. Results from the pinned stack (vLLM ROCm nightly 0.26.1rc1.dev229+g124154a88.rocm723, AITER 0.1.19): Metric Result Single-stream decode (median per-stream, DSpark-7) 168.6 tok/s Prefill with tuned kernels ≈ 7.9–8.5K tok/s (6,988–7,019 tok/s on fresh prompts in the shipping profile) 8 concurrent streams 542 tok/s aggregate, 90.3 tok/s median per stream 64-stream burst 830 tok/s aggregate, no OOM, no engine errors Context 256K validated (the architecture supports 1M) Weights in HBM 156.67 GiB — no additional quantization or weight offload The official vLLM recipe targets NVIDIA and newer AMD hardware. Running the model reliably on MI300X required fixes for its FP8 format, MoE routing at high concurrency, causal speculative verification, CPU-KV synchronization, and several untuned kernel shapes. This repository collects those fixes and pins the versions used in production. Why MI300X The MI300X has 192 GB of HBM3 and 5.3 TB/s of memory bandwidth, with 2.4× the HBM capacity of an H100 SXM5 (AMD). Doubleword's write-up estimates that it costs roughly half as much at list price. For this 304B-parameter checkpoint, the memory capacity allows a simple single-GPU deployment: The entire model fits in HBM without PCIe weight streaming or layer offload. There is room for a 20 GB GPU KV pool and a 96 GiB CPU tier for evicted prefix-cache entries. One card handles 2–8 typical concurrent streams and bursts of up to 64 streams. MI300X (CDNA3) implements the AMD/Graphcore fnuz variant of E4M3, while MI325X and newer use OCP-standard FP8 (background). A kernel that assumes OCP semantics on MI300X can be wrong by a factor of two in the scale domain. Correctness on this FP8 implementation was the first priority; performance tuning came afterward. Prior art, and what this repo adds Fergus Finn's MI300X worklog and the accompanying Doubleword repository identified the FP8 incompatibility, missing AITER fast paths on gfx942, HIP-graph hazards in sparse MLA decode, and MoE routing bugs. The official vLLM recipe covers NVIDIA hardware and newer AMD GPUs (MI325X at 4K context and MI355X), but not a single-MI300X production configuration for the 0731 checkpoint. This repository adds: Correctness overlays for the pinned ROCm nightly, including fixes not yet in upstream vLLM. A validated serving configuration with probabilistic DSpark drafting, block rejection, and static K=7. It uses a 2,048-token scheduler budget and a 1,024-token long-prefill cap to prevent a cold prompt from stalling other streams. AITER GEMM tuning tables for the recurring gfx942 shapes the packaged tables were missing, plus a gfx942 OGS geometry override for the MXFP4 experts. A hybrid KV strategy: 20 GB of fp8_ds_mla GPU cache + 96 GiB native CPU offload, with a load-path fencing fix that upstream issue #47282 documents but PR #47291 never merged. Repository layout . ├── compose.yaml # The production stack (vLLM ROCm + Caddy), digest-pinned ├── Caddyfile.example # Copy to Caddyfile; set hostname, email, and source CIDR ├── vllm-entrypoint.sh # Removes stale CPU-KV mmaps from /dev/shm before start ├── SHA256SUMS # SHA-256 pins for every runtime artifact ├── patches/ │ ├── *.py # Byte-for-byte production overlays (mounted read-only) │ ├── diffs/*.patch # Unified diffs vs. the upstream base revision │ └── README.md # Provenance and regeneration instructions └── tuning/ └── *.csv # AITER A8W8 blockscale tuning tables for gfx942 Runtime configuration The stack uses a digest-pinned official vLLM ROCm nightly with: --trust-remote-code and the DeepSeek V4 tokenizer, reasoning, and tool parsers fp8_ds_mla KV cache (UE8M0 block-scaled FP8, not generic unscaled FP8) with 256-token blocks VLLM_ROCM_USE_AITER=1 and --moe-backend triton; Triton OGS handles the grouped MXFP4 experts, while AITER handles attention and dense linear layers DSpark-7 speculative decoding with probabilistic drafting and block rejection full/breakable CUDA graph capture, giving one graph launch per token during steady decode Caddy as an IP-allowlisted HTTPS proxy Deploying it 1. Host prerequisites One MI300X (gfx942, 304 CUs, ~192 GiB HBM), a working AMD kernel driver, recent Docker Compose, ~235 GiB RAM for the CPU KV tier, and ~500 GB disk (the model cache alone is ~156 GB). 2. Pull the pinned runtime and model VLLM_IMAGE='vllm/vllm-openai-rocm@sha256:e68d18b2ba50298661bfc49baf01158fbf036645c2362cccf3e8a7a79fe6c69a' MODEL='deepseek-ai/DeepSeek-V4-Flash-0731' REVISION='7872f01b1d1fe23eabc4c98b48bffcef5a386062' docker pull "$VLLM_IMAGE" docker run --rm --entrypoint hf \ -v /root/.cache/huggingface:/root/.cache/huggingface \ "$VLLM_IMAGE" download "$MODEL" --revision "$REVISION" 3. Prepare the files cp Caddyfile.example Caddyfile # then set your hostname, email, and remote_ip CIDR mkdir -p aiter-cache crash-dumps chmod +x vllm-entrypoint.sh sha256sum -c SHA256SUMS # verify the overlays before first start 4. Start docker compose config -q docker compose up -d docker compose logs -f inference A healthy start takes ~5 minutes and must show all of: Model loading took 156.67 GiB DSpark draft model loaded: 96 params GPU KV cache size: 1,927,444 tokens Maximum concurrency for 262,144 tokens per request: 7.35x Created mmap file /dev/shm/vllm_offload_...mmap (103.08 GB) Capturing CUDA graphs (FULL) Application startup complete After graph capture, run rocm-smi --showmeminfo vram. The warmed high-water mark is ~204.5 GB of 205.8 GB. If only a few hundred MB remain, the server may start but fail on the first request. 5. Smoke-test HOST='your-host.example.com' curl -fsS "https://$HOST/v1/models" curl -sS "https://$HOST/v1/completions" \ -H 'Content-Type: application/json' \ -d "{\"model\": \"deepseek-ai/DeepSeek-V4-Flash-0731\", \"prompt\": \"Calculate 17 * 23. Answer with the number only.\", \"temperature\": 0, \"max_tokens\": 32}" The patches Each patches/*.py file is a full-file overlay mounted read-only over its counterpart in the container; compose.yaml contains the target paths. The corresponding diffs/*.patch records the change from its upstream base. The base image remains digest-pinned, so upgrades require changing the image reference and revalidating the stack. Overlay Mounted over Fixes Needed when gpt_oss_triton_kernels_moe.pack128-fused-silu-fast-routing.py vllm/.../fused_moe/experts/gpt_oss_triton_kernels_moe.py MXFP4 bitmatrix padding lanes + fused-SiLU grouped experts + fast DeepSeek routing Required for the MXFP4 Triton path; the mask fix is not yet upstream mxfp4.fused-silu.py vllm/.../fused_moe/oracle/mxfp4.py Gate/up interleave layout for the fused-SiLU kernel Required with the fused-SiLU overlay; skip both if you keep the standard SiLU path triton-kernels-matmul-ogs-opt-flags.dsv4-mi300x.py vllm/third_party/triton_kernels/matmul_ogs_details/opt_flags.py gfx942 MXFP4 OGS tile geometry (up to 1,536 routed rows) Performance on gfx942; the stock geometry slows sharply above 768 routed rows fused_compress_quant_cache.fnuz-shuffle.py vllm/models/deepseek_v4/common/ops/fused_compress_quant_cache.py FNUZ FP8 + 16×16 preshuffle in the Lightning Indexer cache writer Required on MI300X; MI325X/MI355X use OCP FP8 and must keep the stock bytes aiter_pa_mqa_logits.i64.py aiter/ops/triton/gluon/pa_mqa_logits.py 64-bit offsets in the ChunkK=256 paged-MQA kernels Required when KV offsets can exceed 4 GiB; skip for small KV pools rocm_aiter_mla_sparse.prefill-bh64.py vllm/v1/attention/ops/rocm_aiter_mla_sparse.py Deterministic torch.topk prefill + BLOCK_H=64 head-512 sparse prefill Determinism is required for reproducible tool calls; BLOCK_H=64 is performance rocm_aiter_mla.dspark-causal.py vllm/v1/attention/backends/mla/rocm_aiter_mla.py Causal multi-token speculative verification Required for DSpark on ROCm small-head MLA — now upstream; the overlay is the upstream file verbatim dspark-speculator.independent-draft-gumbel.py + spec-decode-utils.independent-draft-gumbel.py vllm/v1/worker/gpu/spec_decode/dspark/speculator.py + .../spec_decode/utils.py Draft-proposal Gumbel noise salted away from rejection/recovery noise Required only with draft_sample_method=probabilistic (the recipe's greedy path does not need it) kv_offload_cpu_gpu_worker.load-war.py vllm/v1/kv_offload/cpu/gpu_worker.py Fence CPU→GPU KV restores behind in-flight compute (#47282, PR #47291) Required only with --kv-offloading-backend native Two important correctness fixes MXFP4 routing. The MoE bitmatrix kernel pads its block columns to a Triton block size, but the padding lanes were masked against the global tensor bound instead of the logical block size. Under load, padded lanes corrupted the routing matrix, causing near-match tool names and forgotten schemas on long prompts. The one-line fix is mask = (offs_local < BLOCK_SIZE) & (offs_global < nonzero_indx_size), taken from Doubleword commit c32932bb9. The overlay also includes fused-SiLU and fast-routing changes for grouped MXFP4 experts. FP8 format. DeepSeek V4's Lightning Indexer cache uses FP8. The stock writer emits OCP E4M3 bytes in row-major order, while AITER on MI300X consumes AMD FNUZ E4M3 bytes in a preshuffled 16×16 tile layout. In the worst case, interpreting one format as the other produces a factor-of-two scale error. The overlay selects float8e4b8 with FP8_MAX=224.0 and shuffled write offsets on ROCm, while leaving the OCP path unchanged elsewhere. Speculative decoding This stack uses probabilistic drafting with block rejection. The two Gumbel overlays keep draft-proposal noise independent of rejection and recovery noise. Performance