# vLLM — cau hinh cho beyoru/KAT-Coder-V2.5-Dev-VL-Flash (NVFP4).
#
# ⚠️ HAI CANH BAO, doc truoc khi dung:
#
# 1. BAT BUOC GPU BLACKWELL (sm100/sm120): B200, B300, RTX PRO 6000 Blackwell,
#    RTX 50-series, DGX Spark. NVFP4 chay weight 4-bit VA activation 4-bit tren
#    FP4 tensor core. KHONG chay tren Hopper (H100/H200), Ada, Ampere.
#    Tren nhung card do hay dung ban bf16: beyoru/KAT-Coder-V2.5-Dev-VL
#
# 2. CHUA KIEM CHUNG TREN PHAN CUNG THAT. Checkpoint duoc dung tren Hopper, noi
#    NVFP4 khong chay duoc. Shape weight/scale, quantization_config va cau truc
#    checkpoint da verify TINH; chua sinh mot token nao tren FP4 tensor core that.
services:
  kat-vl-flash:
    image: vllm/vllm-openai:v0.26.0
    container_name: kat-vl-flash
    ipc: host
    ports:
      - "8000:8000"
    volumes:
      - /path/to/models:/models
      - ${HOME}/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ["0"]
              capabilities: [gpu]
    command: >
      --model /models/KAT-Coder-V2.5-Dev-VL-Flash
      --served-model-name kat-vl-flash
      --max-model-len 32768
      --max-num-batched-tokens 8192
      --gpu-memory-utilization 0.90
      --trust-remote-code

# ⚠️ `--max-num-batched-tokens 8192` la BAT BUOC, dung bo di.
#
# Kien truc nay lai attention + Gated DeltaNet. Khi prefix caching bat (mac dinh o
# vLLM moi), vLLM ep `mamba_cache_mode='align'`, roi phai nang attention block size
# len 2096 de "attention page size >= mamba page size". Mac dinh
# max_num_batched_tokens chi 2048 => assert vo ngay luc khoi tao KV cache:
#
#   AssertionError: In Mamba cache align mode,
#                   block_size (2096) must be <= max_num_batched_tokens (2048)
#
# Bat ky gia tri nao >= 2096 deu qua; 8192 con loi cho prefill. Cach khac la tat
# prefix caching (mamba cache ve mode 'none') nhung the thi mat prefix cache — khong dang.

# ⛔ DUNG ghim MoE backend. De vLLM tu chon. Tren Blackwell duong nhanh la
#    CUTLASS / FlashInfer-TRTLLM / CuTe-DSL; ep Marlin se tut manh vi Marlin la
#    kernel weight-only (W4A16), khong bao gio cham toi FP4 tensor core.
#
# Weight chi 22 GB (giam 68.7% so voi 70.2 GB bf16) nen vua thoai mai mot card
# Blackwell don. KV cache da co scale fp8 hieu chuan san trong checkpoint.
#
# So do do chinh xac (doc duoc tu config.json):
#   NVFP4 W4A4 : toan bo mlp.experts.*.{gate,up,down}_proj + shared_expert (group 16)
#   FP8  W8A8  : self_attn.{q,k,v,o}_proj, linear_attn.{in_proj_qkv,in_proj_z,out_proj}, lm_head
#   BF16       : router mlp.gate, shared_expert_gate, linear_attn.{in_proj_a,in_proj_b,norm},
#                va TOAN BO vision tower (333 tensor)
#   KV cache   : fp8 per-tensor static