LordNeel/DeepSeek-V4-Flash-Acti-MTP-W4A16-FP8 at main

156 GB

Ctrl+K

1 contributor

Add measured MTP draft-acceptance metrics (77.7% @128k , 84.9% @262k prod, 93-96% code) + acceptance chart; update manifest

f561982 verified 12 days ago

charts
Add measured MTP draft-acceptance metrics (77.7% @128k, 84.9% @262k prod, 93-96% code) + acceptance chart; update manifest 12 days ago
docker
docker: end-to-end verified — runtime base is cudnn-devel (nvcc) + python3.12-dev + gcc for Triton JIT; image launches vLLM and serves /v1/models + /v1/chat/completions 2 months ago
.gitattributes

1.52 kB
initial commit 2 months ago
AGENTS.md

16.6 kB
Add AGENTS.md — end-to-end runbook for AI agents to set up + serve this model 2 months ago
MTP_QUANT_MANIFEST.json

1.4 kB
Add measured MTP draft-acceptance metrics (77.7% @128k, 84.9% @262k prod, 93-96% code) + acceptance chart; update manifest 12 days ago
MTP_QUANT_MANIFEST.md

1.04 kB
Add measured MTP draft-acceptance metrics (77.7% @128k, 84.9% @262k prod, 93-96% code) + acceptance chart; update manifest 12 days ago
README.md

21.1 kB
Add measured MTP draft-acceptance metrics (77.7% @128k, 84.9% @262k prod, 93-96% code) + acceptance chart; update manifest 12 days ago
config.json

12.9 kB
Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
generation_config.json

174 Bytes
Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
model-00001-of-00004.safetensors

50 GB
xet

Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
model-00002-of-00004.safetensors

50 GB
xet

Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
model-00003-of-00004.safetensors

50 GB
xet

Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
model-00004-of-00004.safetensors

2.48 GB
xet

Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
model-mtp-w4a16.safetensors

3.55 GB
xet

v2: real GPTQ on MTP routed experts (replaces v1 RTN). Added MTP_QUANT_MANIFEST is_final_quality_preserving=true. 2 months ago
model.safetensors.index.json

8.68 MB
v2: real GPTQ on MTP routed experts (replaces v1 RTN). Added MTP_QUANT_MANIFEST is_final_quality_preserving=true. 2 months ago
recipe.yaml

1.97 kB
Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
tokenizer.json

10.1 MB
Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago
tokenizer_config.json

397 Bytes
Initial release: pasta-paul base + Acti MTP block (RTN INT4 routed experts, FP8_BLOCK attn) 2 months ago