# Prefill-first variant of configs/agent-128k.env for 2x RTX 3090 + 128 GiB RAM: 128K context, hot88,
# 8,192-token prefill chunks. Copy these values into .env instead of the agent profile.
# Every prefill chunk streams all cold experts once, so 8,192-token chunks halve that traffic on long prompts
# (about 30% faster prefill at 131K). The VRAM they need comes from 12 fewer hot experts per GPU, so each
# decode step pulls more experts from system memory (10-14% slower decode). With hot100, 8,192-token chunks
# run out of VRAM. Single request, up to 131,072 prompt tokens plus 4,096 generated tokens.

MAX_MODEL_LEN=135168
# 110 KV blocks: a full 135,168-token request plus the extra recurrent-state
# block per Mamba group that async prefill holds (reported concurrency 1.04x).
KV_CACHE_MEMORY_BYTES=2500000000
VLLM_WNA16_STATIC_HOT_CACHE_SIZE=88
MAX_NUM_BATCHED_TOKENS=8192

# Runtime switches, as in configs/fast-256k.env.
QWEN38_STREAM_STAGE=1
QWEN38_STREAM_STAGE_MIN_TOKENS=1536
QWEN38_STAGE_OVERLAP=1
QWEN38_ASYNC_SCHEDULING=1
QWEN38_TRITON_SKINNY=1
VLLM_MTP_DRAFT_VOCAB_RANGES=[[0,65536],[248044,248320]]
DISABLE_CUSTOM_ALL_REDUCE=0
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:False
QWEN38_PLE_PREFAULT=1
JIT_CACHE_DIR=./jit-cache