memra-engine 0.83.0

From-scratch CUDA LLM inference engine for NVIDIA RTX 50-series (sm_120a) and Hopper (sm_90a) - custom kernels, no frameworks
Documentation
[package]
name = "memra-engine"
description = "From-scratch CUDA LLM inference engine for NVIDIA RTX 50-series (sm_120a) and Hopper (sm_90a) - custom kernels, no frameworks"
version.workspace = true
license.workspace = true
repository.workspace = true
rust-version.workspace = true
edition = "2024"

[dependencies]
memra-sampling = { workspace = true }
memra-validate = { workspace = true }
memra-kv = { workspace = true }
memra-gguf = { workspace = true }
memra-runtime = { workspace = true }
memra-tokenizer = { workspace = true }
cudarc = { version = "0.19", default-features = false, features = ["std", "cuda-13010", "dynamic-loading", "driver", "cublas", "cublaslt", "nvrtc"] }
memmap2 = "0.9"
libc = "0.2"

[[bin]]
name = "gemma-gate"
path = "src/bin/gemma_gate.rs"

[[bin]]
name = "prime-graph-gate"
path = "src/bin/prime_graph_gate.rs"

[[bin]]
name = "prime-graph-smoke"
path = "src/bin/prime_graph_smoke.rs"

[[bin]]
name = "prime-batch-gate"
path = "src/bin/prime_batch_gate.rs"

[[bin]]
name = "prime-gate"
path = "src/bin/prime_gate.rs"

[[bin]]
name = "concat-prime-probe"
path = "src/bin/concat_prime_probe.rs"

[[bin]]
name = "iso-gap-probe"
path = "src/bin/iso_gap_probe.rs"

[[bin]]
name = "k27div-probe"
path = "src/bin/k27div_probe.rs"

[[bin]]
name = "argmax-margin-probe"
path = "src/bin/argmax_margin_probe.rs"

[[bin]]
name = "kernel-check"
path = "src/bin/kernel_check.rs"

[[bin]]
name = "router-batch-bench"
path = "src/bin/router_batch_bench.rs"

[[bin]]
name = "mvq-msweep"
path = "src/bin/mvq_msweep.rs"

[[bin]]
name = "fa-sanitize"
path = "src/bin/fa_sanitize.rs"

[[bin]]
name = "kv-fmt-bench"
path = "src/bin/kv_fmt_bench.rs"

[[bin]]
name = "fa-hd128-check"
path = "src/bin/fa_hd128_check.rs"

[[bin]]
name = "quant-kv-f32-check"
path = "src/bin/quant_kv_f32_check.rs"

[[bin]]
name = "run-dense"
path = "src/bin/run_dense.rs"

[[bin]]
name = "run-hybrid"
path = "src/bin/run_hybrid.rs"

[[bin]]
name = "run-safetensors"
path = "src/bin/run_safetensors.rs"

[[bin]]
name = "st-vs-gguf"
path = "src/bin/st_vs_gguf.rs"

[[bin]]
name = "frspec-owngen"
path = "src/bin/frspec_owngen.rs"

[[bin]]
name = "dspark-generate"
path = "src/bin/dspark_generate.rs"

[[bin]]
name = "dspark-extract"
path = "src/bin/dspark_extract.rs"

[[bin]]
name = "dspark-export-shared"
path = "src/bin/dspark_export_shared.rs"

[[bin]]
name = "run-gen"
path = "src/bin/run_gen.rs"

[[bin]]
name = "run-spec"
path = "src/bin/run_spec.rs"

[[bin]]
name = "spec-econ"
path = "src/bin/spec_econ.rs"

[[bin]]
name = "replay-acceptance"
path = "src/bin/replay_acceptance.rs"

[[bin]]
name = "run-eagle"
path = "src/bin/run_eagle.rs"


[[bin]]
name = "decode-bench"
path = "src/bin/decode_bench.rs"

[[bin]]
name = "argmax-gate"
path = "src/bin/argmax_gate.rs"

[[bin]]
name = "w4a4-gate"
path = "src/bin/w4a4_gate.rs"

[[bin]]
name = "moe-devq8-check"
path = "src/bin/moe_devq8_check.rs"

[[bin]]
name = "decode-dc-gate"
path = "src/bin/decode_dc_gate.rs"

[[bin]]
name = "graph-decode-gate"
path = "src/bin/graph_decode_gate.rs"

[[bin]]
name = "cutlass-smoke"
path = "src/bin/cutlass_smoke.rs"

[[bin]]
name = "gdn-bench"
path = "src/bin/gdn_bench.rs"

[[bin]]
name = "verify-probe"
path = "src/bin/verify_probe.rs"

[[bin]]
name = "verify-mscale"
path = "src/bin/verify_mscale.rs"

[[bin]]
name = "depth-profile"
path = "src/bin/depth_profile.rs"

[[bin]]
name = "decode-window-profile"
path = "src/bin/decode_window_profile.rs"

[[bin]]
name = "session-gate"
path = "src/bin/session_gate.rs"

[[bin]]
name = "spec-serve-gate"
path = "src/bin/spec_serve_gate.rs"

[[bin]]
name = "optipipe-gate"
path = "src/bin/optipipe_gate.rs"

[[bin]]
name = "target-efficiency"
path = "src/bin/target_efficiency.rs"

[[bin]]
name = "session-bench"
path = "src/bin/session_bench.rs"

[[bin]]
name = "fa-ab-bench"
path = "src/bin/fa_ab_bench.rs"

[[bin]]
name = "frspec-rank"
path = "src/bin/frspec_rank.rs"

[[bin]]
name = "decode-batch-gate"
path = "src/bin/decode_batch_gate.rs"

[[bin]]
name = "decode-batch-bench"
path = "src/bin/decode_batch_bench.rs"

[[bin]]
name = "graph-session-gate"
path = "src/bin/graph_session_gate.rs"

[[bin]]
name = "graph-allocfree-probe"
path = "src/bin/graph_allocfree_probe.rs"

[[bin]]
name = "graph-warmup-stress"
path = "src/bin/graph_warmup_stress.rs"

[[bin]]
name = "pp2-gate"
path = "src/bin/pp2_gate.rs"

[[bin]]
name = "pp-transport-smoke"
path = "src/bin/pp_transport_smoke.rs"

[[bin]]
name = "ppn-gate"
path = "src/bin/ppn_gate.rs"

[[bin]]
name = "ppn-bench"
path = "src/bin/ppn_bench.rs"

[[bin]]
name = "clc-mmq-bench"
path = "src/bin/clc_mmq_bench.rs"

[[bin]]
name = "spec-st-probe"
path = "src/bin/spec_st_probe.rs"

[[bin]]
name = "plainbatch-margin-probe"
path = "src/bin/plainbatch_margin_probe.rs"

[[bin]]
name = "accprobe-bench"
path = "src/bin/accprobe_bench.rs"

[[bin]]
name = "gemv-e4m3-bench"
path = "src/bin/gemv_e4m3_bench.rs"