[Doc] Document ECMooncakeConnector for EPD disaggregation - #2

Draft

stmatengss wants to merge 20 commits into

mainfrom

stmatengss/docs-epd-mooncake-e48d

Commits on Sep 14, 2026

[Refactor] Normalize DeepSeek V4.1 model package naming (#56741)
BugenZhao
authored
cf1584f
[Fast Start] Use GPU uuid as socket folder identifier (#56669)
Isotr0py
authored
dbf49da
[Bugfix] Fix --lora-modules name=path parsing when path contains '=' (#53353)
hungnnvidia
authored
663d7f6
[Bugfix][Rust Frontend] Preserve selected-token logprob mode (#56406)

alec-flowers
and
codex
authored
e6b4e47
[Performance] Add Triton kernel for Gemma3n sparse GELU (#48498)

authored
78e8426
[CI] Update entrypoints CI (#56763)
noooop
authored
23cfaad
[Perf][GLM-5.3-Flash] Use FlashKDA for KDA chunked prefill (1.7-3.8x faster than the Triton chunk path) (#55737)

authored
b443c1c
[CPU][Profiler] Group torch profiler tables by input shape when record_shapes is on (#56016)

Chinmay-Kulkarni-AMD
and
bigPYJ1151
authored
dc89fdf
[Bugfix][CPU] Fix DeepSeek-R1 (FP8 MLA + MoE) correctness on CPU backend (#56773)
bigPYJ1151
authored
79f0be2
[Perf][GLM-5.3-Flash] Dense/masked-MHA sparse prefill for the NoPE (256, 0, 256) layout + skip the NoPE K concat (#55738)

authored
238cb2b
[ROCm][Perf] W4A16: keep skinny GEMM zero-points packed 4-bit (#54965)
mgehre-amd
authored
dc2e8f1
[CI] Keep OTel bytecode out of mounted checkouts (#56766)

khluu
and
codex
authored
5236bef
[Bugfix][Mooncake] Report failed remote KV loads to the scheduler (#50984)

authored
2c7ee87
[Frontend] Replace `VLLM_ENABLE_SCALE_OUT_ENDPOINTS` with `--enable-scale-out` (#55176)
hickeyma
authored
934b1fc
[Bugfix][EPD] Preserve media processing options in encoder requests (#56786)
jiangkuaixue123
authored
1678b39
[Security] Cap Qwen-VL video sampling knobs (#56729)

jperezdealgaba
and
cursoragent
authored
ea723c8
[Bugfix][NixlPush] Guard _remote_agents read in _do_send_reg_notif (X1) (#56317)
NickLucche
authored
0ca7aef
[Bugfix][Rust Frontend][Multimodal] Align DeepSeek V4.1 and Kimi K3 media with rendered placeholders (#56366)

authored
7b1ea3f
[CI][Test] Mock CPU backend block sizes in kv_connector unit conftest (#56671)

xiaozhenbi
and
claude
authored
e82794f
[Doc] Document ECMooncakeConnector for EPD disaggregation
stmatengss
committed
3d66f02

Navigation Menu

[Doc] Document ECMooncakeConnector for EPD disaggregation - #2

[Doc] Document ECMooncakeConnector for EPD disaggregation#2
stmatengss wants to merge 20 commits into
mainfrom
stmatengss/docs-epd-mooncake-e48d

Commits on Sep 14, 2026

[Refactor] Normalize DeepSeek V4.1 model package naming (#56741)

[Fast Start] Use GPU uuid as socket folder identifier (#56669)

[Bugfix] Fix --lora-modules name=path parsing when path contains '=' (#53353)

[Bugfix][Rust Frontend] Preserve selected-token logprob mode (#56406)

[Performance] Add Triton kernel for Gemma3n sparse GELU (#48498)

[CI] Update entrypoints CI (#56763)

[Perf][GLM-5.3-Flash] Use FlashKDA for KDA chunked prefill (1.7-3.8x faster than the Triton chunk path) (#55737)

[CPU][Profiler] Group torch profiler tables by input shape when record_shapes is on (#56016)

[Bugfix][CPU] Fix DeepSeek-R1 (FP8 MLA + MoE) correctness on CPU backend (#56773)

[Perf][GLM-5.3-Flash] Dense/masked-MHA sparse prefill for the NoPE (256, 0, 256) layout + skip the NoPE K concat (#55738)

[ROCm][Perf] W4A16: keep skinny GEMM zero-points packed 4-bit (#54965)

[CI] Keep OTel bytecode out of mounted checkouts (#56766)

[Bugfix][Mooncake] Report failed remote KV loads to the scheduler (#50984)

[Frontend] Replace `VLLM_ENABLE_SCALE_OUT_ENDPOINTS` with `--enable-scale-out` (#55176)

[Bugfix][EPD] Preserve media processing options in encoder requests (#56786)

[Security] Cap Qwen-VL video sampling knobs (#56729)

[Bugfix][NixlPush] Guard _remote_agents read in _do_send_reg_notif (X1) (#56317)

[Bugfix][Rust Frontend][Multimodal] Align DeepSeek V4.1 and Kimi K3 media with rendered placeholders (#56366)

[CI][Test] Mock CPU backend block sizes in kv_connector unit conftest (#56671)

[Doc] Document ECMooncakeConnector for EPD disaggregation