mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 14:17:52 +00:00
The tree names only Hanzo Engine, and these files cannot be reworded into
that: what they do is run, wrap or report on another inference engine.
- releases/v0.8.2: a release report imported with the fork's history. Every
table and chart is a speed ratio against two other engines, and nothing
links to it.
- gpu/: Windows PowerShell that downloads another engine's binaries and
serves them under our name. Nothing calls it.
- scripts that drive another engine as a baseline or as the recorder of a
golden: the head-to-head bench and perplexity gate, the block-FP8 golden
and bench, the MoE and zen6 vector recorders, the zen6 MoE golden, dump,
bench and window scripts. The fixtures they recorded stay where the tests
read them; the recorders remain in history at 87d90f4d08.
examples/server/phi3_duckduckgo_mistral.rs.ipynb becomes
phi3_duckduckgo_hanzo.ipynb, the name docs/EXAMPLES.md and .typos.toml
already link to, with its recorded outputs cleared (they were a web search
about the old name).
Co-authored-by: Hanzo Dev <[email protected]>
96 lines
2.5 KiB
TOML
96 lines
2.5 KiB
TOML
[files]
|
|
extend-exclude = [
|
|
".git/",
|
|
"calibration_data/",
|
|
"examples/server/phi3_duckduckgo_hanzo.ipynb",
|
|
"hanzo-cli/static/",
|
|
"CLAUDE.md",
|
|
"hanzo-web-chat/static/",
|
|
]
|
|
ignore-hidden = false
|
|
|
|
[default]
|
|
extend-ignore-re = [
|
|
"cudaDevAttrMaxSharedMemoryPerBlockOptin",
|
|
'"tese"',
|
|
"setp\\.ne\\.b32",
|
|
# metal_kernels
|
|
"out uneeded values",
|
|
"copy_gg_dynamic_nd1",
|
|
"copy_gg_dynamic_nd2",
|
|
"copy_gg_dynamic_nd3",
|
|
"copy_g_nd1",
|
|
"copy_g_nd2",
|
|
"copy_g_nd3",
|
|
"copy_gg_nd1",
|
|
"copy_gg_nd2",
|
|
"copy_gg_nd3",
|
|
"elem_to_loc_2_nd",
|
|
"elem_to_loc_3_nd",
|
|
"thr_reduce",
|
|
"(?i)thw",
|
|
"iy",
|
|
"mxfp4_get_max_smem_optin",
|
|
"alog",
|
|
"ba",
|
|
# mmq_gguf CUDA kernels (ggml port)
|
|
"ggml_cuda_get_max_cpy_bytes",
|
|
"nb_per_cpy",
|
|
"ggml_cuda_ue4m3_to_fp32",
|
|
"ue4m3",
|
|
"UE4M3",
|
|
"ue8m0",
|
|
"UE8M0",
|
|
"CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN",
|
|
"PNGs",
|
|
]
|
|
|
|
[default.extend-words]
|
|
"ammount" = "ammount"
|
|
# ROCm Heterogeneous System Architecture (HSA_XNACK env var)
|
|
"HSA" = "HSA"
|
|
# least-squares accumulator sum(w * x * Laux) in convert_fp8_to_gguf.py
|
|
"sxl" = "sxl"
|
|
# TRELLIS flexicubes interpolated-edge variable
|
|
"ue" = "ue"
|
|
# 3D-generation route/module naming (/v1/3d)
|
|
"threed" = "threed"
|
|
"arange" = "arange"
|
|
"layrnorm" = "layrnorm"
|
|
"ot" = "ot"
|
|
"thr" = "thr"
|
|
"mis" = "mis"
|
|
"Hel" = "Hel"
|
|
"nothink" = "nothink"
|
|
|
|
[type.rust.extend-identifiers]
|
|
# upstream MultiTalk class name is spelled this way
|
|
"SingleStreamMutiAttention" = "SingleStreamMutiAttention"
|
|
"arange" = "arange"
|
|
"arange_step" = "arange_step"
|
|
"cna" = "cna"
|
|
"from_mmaped_safetensors" = "from_mmaped_safetensors"
|
|
"MmapedSafetensors" = "MmapedSafetensors"
|
|
"nin_shortcut" = "nin_shortcut"
|
|
"EmbedNd" = "EmbedNd"
|
|
"gather_nd" = "gather_nd"
|
|
"i_thw" = "i_thw"
|
|
"cached_img_thw" = "cached_img_thw"
|
|
"cached_vid_thw" = "cached_vid_thw"
|
|
"video_grid_thw" = "video_grid_thw"
|
|
"image_grid_thw" = "image_grid_thw"
|
|
"grid_thw" = "grid_thw"
|
|
"vision_grid_thw" = "vision_grid_thw"
|
|
"video_grid_thw_accum" = "video_grid_thw_accum"
|
|
"image_grid_thw_accum" = "image_grid_thw_accum"
|
|
"image_thw" = "image_thw"
|
|
"video_thw" = "video_thw"
|
|
"depthwise_seperable_out_channel" = "depthwise_seperable_out_channel"
|
|
"default_depthwise_seperable_out_channel" = "default_depthwise_seperable_out_channel"
|
|
"DepthWiseSeperableConv1d" = "DepthWiseSeperableConv1d"
|
|
"layrnorm" = "layrnorm"
|
|
"pre_layrnorm" = "pre_layrnorm"
|
|
"numer_denom_ms" = "numer_denom_ms"
|
|
"temperal_downsample" = "temperal_downsample"
|
|
"temperal_up" = "temperal_up"
|