vLLM-Moet
A vLLM patch + hand‑written SM120 SASS kernels: 2‑bit MoE experts + an FP4 "delta" cache that recovers precision — matching the official (NV)FP4 checkpoint's quality on consumer Blackwell cards
File Explorer
- bench-lint.yml
- native-tp2-c1-32k-mtp1-gpqa.json
- native-tp2-c1-32k-mtp1-gsm8k.json
- native-tp2-c1-32k-mtp2g-gpqa.json
- native-tp2-c1-32k-mtp2g-gsm8k.json
- native-tp2-gpqa-nt-0731.json
- native-tp2-gpqa-nt-v2.json
- native-tp2-gpqa-nt.json
- native-tp2-gpqa-think-high-v2.json
- native-tp2-gpqa-think-high.json
- native-tp2-gsm8k-c1.json
- native-tp2-gsm8k-c2-0731.json
- native-tp2-gsm8k-c2-v2.json
- native-tp2-gsm8k-c2.json
- registry.yaml
- fin03-pro6000x2.yaml
- rtx-5090x4.yaml
- rtx-pro6000x4.yaml
- rtx-pro6000x8.yaml
- runpod-4xpro6000-docker.yaml
- runpod-4xpro6000.yaml
- 5090x1-basecache.yaml
- 5090x1-exact.yaml
- 5090x2-exact.yaml
- 5090x2-tp2-basecache.yaml
- 5090x4-tp4.yaml
- pro6000x1-cache.yaml
- pro6000x1.yaml
- pro6000x2-tp2-maxq.yaml
- pro6000x2-tp2-native-c1-mtp1.yaml
- pro6000x2-tp2-native-c1-mtp2g.yaml
- pro6000x2-tp2-native.yaml
- pro6000x2-tp2.yaml
- 5090x2-tp2-basecache.yaml
- 5090x4-tp4.yaml
- pro6000x1.yaml
- pro6000x2-tp2-native.yaml
- pro6000x2-tp2.yaml
- pro6000x2-basecache.yaml
- pro6000x4-tp4-dcp4-1m.yaml
- pro6000x4-tp4-mtp.yaml
- pro6000x4-tp4-quality-mtpk5.yaml
- pro6000x4-tp4-quality-mtpk7.yaml
- pro6000x4-tp4-quality.yaml
- pro6000x2-basecache.yaml
- pro6000x4-tp4-256k.yaml
- pro6000x4-tp4-eagle3.yaml
- deepseek-v4-flash__5090x1-basecache.json
- deepseek-v4-flash__5090x4-tp4.json
- deepseek-v4-flash__pro6000x1-24k.json
- deepseek-v4-flash__pro6000x1-512k.json
- deepseek-v4-flash__pro6000x2-tp2.json
- glm-5.2-nvfp4__pro6000x2-basecache.json
- glm-5.2-nvfp4__pro6000x4-tp4-mtp.json
- glm-5.2-nvfp4__pro6000x4-tp4-quality.json
- kimi-k2.7-code-nvfp4__pro6000x2-basecache.json
- kimi-k2.7-code-nvfp4__pro6000x4-tp4-256k.json
- kimi-k2.7-code-nvfp4__pro6000x4-tp4-eagle3.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gpqa-think.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gpqa.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gsm8k-run2.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gsm8k.json
- deepseek-v4-flash__pro6000x2-tp2-maxq.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gpqa-think.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gpqa.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gsm8k.json
- deepseek-v4-flash__pro6000x2-tp2-maxq.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gpqa.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gsm8k.json
- deepseek-v4-flash__pro6000x2-tp2-maxq.json
- quality__deepseek-v4-flash__pro6000x1-cache__gpqa.json
- quality__deepseek-v4-flash__pro6000x1-cache__gsm8k.json
- quality__deepseek-v4-flash__pro6000x1__gpqa.json
- quality__deepseek-v4-flash__pro6000x1__gsm8k.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gpqa.json
- quality__deepseek-v4-flash__pro6000x2-tp2-maxq__gsm8k.json
- deepseek-v4-flash__pro6000x1-cache.json
- deepseek-v4-flash__pro6000x1.json
- deepseek-v4-flash__pro6000x2-tp2-maxq.json
- bench.py
- common.py
- envinfo.py
- integrity.py
- lint.py
- probes.py
- render.py
- serverctl.py
- sweep.py
- mtp-ab.yaml
- quality-native.yaml
- quality.yaml
- quick.yaml
- standard.yaml
- test_integrity.py
- matrix.yaml
- models.yaml
- README.md
- serve_recipe.py
- baseline-2026-07-10.md
- dcp-collectives-2026-07-13.md
- mtp-acceptance-depth-2026-07-13.md
- v2026.07.17-quality.md
- v2026.07.27-quality.md
- v2026.07.30-quality.md
- v2026.07.31-quality.md
- kimi-k27-code.md
- quality.md
- sm120-alu-subunits.md
- v024-port.md
- dn3_m8.cubin
- dn3_m8g.cubin
- k1_m8.cubin
- k1_m8g.cubin
- ucb2_m8.cubin
- ucb2_m8g.cubin
- uni_m8.cubin
- uni_m8g.cubin
- mla_prefill_state.cubin
- moe_w2_mm_k1024.cubin
- moe_w2_mm_k1024_a32.cubin
- moe_w2_mm_k2048.cubin
- moe_w2_mm_k2048_a32.cubin
- moe_w2_mm_k3584_a32.cubin
- moe_w2_mm_k384.cubin
- moe_w2_mm_k384_a32.cubin
- moe_w2_mm_k4096.cubin
- moe_w2_mm_k4096_a32.cubin
- moe_w2_mm_k512.cubin
- moe_w2_mm_k512_a32.cubin
- moe_w2_mm_k6144.cubin
- moe_w2_mm_k6144_a32.cubin
- moe_w2_mm_k7168.cubin
- moe_w2_mm_k7168_a32.cubin
- moe_w2_mm_k768_a32.cubin
- moe_w2_mm_mc2_k1024.cubin
- moe_w2_mm_mc2_k2048.cubin
- moe_w2_mm_mc2_k384.cubin
- moe_w2_mm_mc2_k4096.cubin
- moe_w2_mm_mc2_k512.cubin
- moe_w2_mm_mc2_k6144.cubin
- moe_w2_mm_mc2_k7168.cubin
- moe_w2_mm_mc4_k1024.cubin
- moe_w2_mm_mc4_k1024_a32.cubin
- moe_w2_mm_mc4_k2048.cubin
- moe_w2_mm_mc4_k2048_a32.cubin
- moe_w2_mm_mc4_k3584_a32.cubin
- moe_w2_mm_mc4_k384.cubin
- moe_w2_mm_mc4_k384_a32.cubin
- moe_w2_mm_mc4_k4096.cubin
- moe_w2_mm_mc4_k4096_a32.cubin
- moe_w2_mm_mc4_k512.cubin
- moe_w2_mm_mc4_k512_a32.cubin
- moe_w2_mm_mc4_k6144.cubin
- moe_w2_mm_mc4_k6144_a32.cubin
- moe_w2_mm_mc4_k7168.cubin
- moe_w2_mm_mc4_k7168_a32.cubin
- moe_w2_mm_mc4_k768_a32.cubin
- moe_w2_mm_mc4afrag_k1024.cubin
- moe_w2_mm_mc4afrag_k1024_a32.cubin
- moe_w2_mm_mc4afrag_k2048.cubin
- moe_w2_mm_mc4afrag_k2048_a32.cubin
- moe_w2_mm_mc4afrag_k3584_a32.cubin
- moe_w2_mm_mc4afrag_k384.cubin
- moe_w2_mm_mc4afrag_k384_a32.cubin
- moe_w2_mm_mc4afrag_k4096.cubin
- moe_w2_mm_mc4afrag_k4096_a32.cubin
- moe_w2_mm_mc4afrag_k512.cubin
- moe_w2_mm_mc4afrag_k512_a32.cubin
- moe_w2_mm_mc4afrag_k6144.cubin
- moe_w2_mm_mc4afrag_k6144_a32.cubin
- moe_w2_mm_mc4afrag_k7168.cubin
- moe_w2_mm_mc4afrag_k7168_a32.cubin
- moe_w2_mm_mc4afrag_k768_a32.cubin
- moe_w4_mm_k1024.cubin
- moe_w4_mm_k1024_a32.cubin
- moe_w4_mm_k2048.cubin
- moe_w4_mm_k2048_a32.cubin
- moe_w4_mm_k3584_a32.cubin
- moe_w4_mm_k384.cubin
- moe_w4_mm_k384_a32.cubin
- moe_w4_mm_k4096.cubin
- moe_w4_mm_k4096_a32.cubin
- moe_w4_mm_k512.cubin
- moe_w4_mm_k512_a32.cubin
- moe_w4_mm_k6144.cubin
- moe_w4_mm_k6144_a32.cubin
- moe_w4_mm_k7168.cubin
- moe_w4_mm_k7168_a32.cubin
- moe_w4_mm_k768_a32.cubin
- moe_w4_mm_k8192.cubin
- moe_w4_mm_k8192_a32.cubin
- moe_w4q_mm_k1024.cubin
- moe_w4q_mm_k1024_a32.cubin
- moe_w4q_mm_k2048.cubin
- moe_w4q_mm_k2048_a32.cubin
- moe_w4q_mm_k4096.cubin
- moe_w4q_mm_k4096_a32.cubin
- moe_w4q_mm_k512.cubin
- moe_w4q_mm_k512_a32.cubin
- moe_w4q_mm_k6144.cubin
- moe_w4q_mm_k6144_a32.cubin
- moe_w4q_mm_k7168.cubin
- moe_w4q_mm_k7168_a32.cubin
- moe_w4s_mm_k1024.cubin
- moe_w4s_mm_k2048.cubin
- moe_w4s_mm_k4096.cubin
- moe_w4s_mm_k512.cubin
- moe_w4s_mm_k6144.cubin
- moe_w4s_mm_k7168.cubin
- culaunch.py
- gen_m8_guard.py
- gen_moe_sfa_probe.py
- gen_moe_w2.py
- gen_moe_w4.py
- gen_moe_w4q.py
- gen_moe_w4q_probe.py
- gen_moe_w4s.py
- moe_sfa_probe_check.py
- moe_w2_afrag_check.py
- moe_w2_check.py
- moe_w2_prefill_bench.py
- moe_w4_check.py
- moe_w4q_check.py
- moe_w4q_probe_check.py
- moe_w4s_check.py
- moe_w4s_nesting_study.py
- dn3_m8.sass
- dn3_m8g.sass
- k1_m8.sass
- k1_m8g.sass
- ucb2_m8.sass
- ucb2_m8g.sass
- uni_m8.sass
- uni_m8g.sass
- mla_prefill_state2.sass
- moe_w2_mm.sass
- moe_w2_mm_k1024.sass
- moe_w2_mm_k1024_a32.sass
- moe_w2_mm_k2048.sass
- moe_w2_mm_k2048_a32.sass
- moe_w2_mm_k3584_a32.sass
- moe_w2_mm_k384.sass
- moe_w2_mm_k384_a32.sass
- moe_w2_mm_k4096_a32.sass
- moe_w2_mm_k512.sass
- moe_w2_mm_k512_a32.sass
- moe_w2_mm_k6144.sass
- moe_w2_mm_k6144_a32.sass
- moe_w2_mm_k7168.sass
- moe_w2_mm_k7168_a32.sass
- moe_w2_mm_k768_a32.sass
- moe_w2_mm_mc2.sass
- moe_w2_mm_mc2_k1024.sass
- moe_w2_mm_mc2_k2048.sass
- moe_w2_mm_mc2_k384.sass
- moe_w2_mm_mc2_k512.sass
- moe_w2_mm_mc2_k6144.sass
- moe_w2_mm_mc2_k7168.sass
- moe_w2_mm_mc4_k1024.sass
- moe_w2_mm_mc4_k1024_a32.sass
- moe_w2_mm_mc4_k2048_a32.sass
- moe_w2_mm_mc4_k3584_a32.sass
- moe_w2_mm_mc4_k384.sass
- moe_w2_mm_mc4_k384_a32.sass
- moe_w2_mm_mc4_k4096_a32.sass
- moe_w2_mm_mc4_k512.sass
- moe_w2_mm_mc4_k512_a32.sass
- moe_w2_mm_mc4_k6144.sass
- moe_w2_mm_mc4_k6144_a32.sass
- moe_w2_mm_mc4_k7168.sass
- moe_w2_mm_mc4_k7168_a32.sass
- moe_w2_mm_mc4_k768_a32.sass
- moe_w2_mm_mc4afrag_k1024.sass
- moe_w2_mm_mc4afrag_k1024_a32.sass
- moe_w2_mm_mc4afrag_k2048.sass
- moe_w2_mm_mc4afrag_k2048_a32.sass
- moe_w2_mm_mc4afrag_k3584_a32.sass
- moe_w2_mm_mc4afrag_k384.sass
- moe_w2_mm_mc4afrag_k384_a32.sass
- moe_w2_mm_mc4afrag_k4096.sass
- moe_w2_mm_mc4afrag_k4096_a32.sass
- moe_w2_mm_mc4afrag_k512.sass
- moe_w2_mm_mc4afrag_k512_a32.sass
- moe_w2_mm_mc4afrag_k6144.sass
- moe_w2_mm_mc4afrag_k6144_a32.sass
- moe_w2_mm_mc4afrag_k7168.sass
- moe_w2_mm_mc4afrag_k7168_a32.sass
- moe_w2_mm_mc4afrag_k768_a32.sass
- moe_w4_mm.sass
- moe_w4_mm_k1024.sass
- moe_w4_mm_k1024_a32.sass
- moe_w4_mm_k2048.sass
- moe_w4_mm_k2048_a32.sass
- moe_w4_mm_k3584_a32.sass
- moe_w4_mm_k384.sass
- moe_w4_mm_k384_a32.sass
- moe_w4_mm_k4096_a32.sass
- moe_w4_mm_k512.sass
- moe_w4_mm_k512_a32.sass
- moe_w4_mm_k6144.sass
- moe_w4_mm_k6144_a32.sass
- moe_w4_mm_k7168.sass
- moe_w4_mm_k7168_a32.sass
- moe_w4_mm_k768_a32.sass
- moe_w4_mm_k8192.sass
- moe_w4_mm_k8192_a32.sass
- moe_w4q_mm_k1024.sass
- moe_w4q_mm_k1024_a32.sass
- moe_w4q_mm_k2048.sass
- moe_w4q_mm_k2048_a32.sass
- moe_w4q_mm_k4096.sass
- moe_w4q_mm_k4096_a32.sass
- moe_w4q_mm_k512.sass
- moe_w4q_mm_k512_a32.sass
- moe_w4q_mm_k6144.sass
- moe_w4q_mm_k6144_a32.sass
- moe_w4q_mm_k7168.sass
- moe_w4q_mm_k7168_a32.sass
- moe_w4s_mm_k1024.sass
- moe_w4s_mm_k2048.sass
- moe_w4s_mm_k4096.sass
- moe_w4s_mm_k512.sass
- moe_w4s_mm_k6144.sass
- moe_w4s_mm_k7168.sass
- qmma_e4m3.merc.stub
- MANIFEST.md
- FILES.txt
- SOURCE.txt
- vllm-moet-v0.24.0.patch
- boot_dcp1m.sh
- COORDINATION.md
- decode_2prompts.py
- decode_at_depth.py
- decode_at_depth_real.py
- run_phase0.sh
- test_dcp_kernels_gpu.py
- test_dcp_lse_merge_sm120.py
- test_dcp_sparse_topk.py
- bench_tok.py
- check_patch_files.py
- decode_ab.py
- delta_sim.py
- delta_trace_demo.py
- import_quality_results.py
- import_tau060_quality.py
- needle_full.py
- needle_probe.py
- prefill_probe.py
- probe_quant_quality.py
- repack_expert_bits.py
- stamp_pack_identity.py
- sweep_nvfp4_codebook.py
- test_bf12.py
- test_moe_w2_base_delta_tp2.py
- test_moe_w2_base_verify.py
- test_moe_w2_forward.py
- test_moe_w2_planes.py
- test_moe_w2_prefill_fp4.py
- test_mqa_logits.py
- test_resident_split_storm.py
- test_split_equiv.py
- test_store_backends.py
- test_three_tier_split.py
- test_three_tier_stress.py
- .dockerignore
- .gitignore
- AGENTS.md
- Dockerfile.recipes
- Dockerfile.sm120-v024
- LICENSE
- NOTICE
- README.md
# Use via CDN
jsDelivrjsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.
Command Glossary
Commands referenced in this DOCs, explained below.
docker run
View Details ▼
docker run
This command is an alias of `docker container run`.
tldr docker container run
View documentation for the original command:
git clone
View Details ▼
git clone
Clone an existing repository.
git clone {{remote_repository_location}} {{path/to/directory}}
Clone an existing repository into a new directory (the default directory is the repository name):
git clone --recursive {{remote_repository_location}}
Clone an existing repository and its submodules:
git clone {{[-n|--no-checkout]}} {{remote_repository_location}}
Clone only the `.git` directory of an existing repository:
sass
View Details ▼
sass
Convert SCSS or Sass files to CSS.
sass {{inputfile.scss|inputfile.sass}}
Convert a SCSS or Sass file to CSS and print out the result:
sass {{inputfile.scss|inputfile.sass}} {{outputfile.css}}
Convert a SCSS or Sass file to CSS and save the result to a file:
sass --watch {{inputfile.scss|inputfile.sass}}
Watch a SCSS or Sass file for changes and output or update the CSS file with same filename:
