attention: support quantized KV cache in MLA DCP context prefill - #14

Closed

voipmonitor wants to merge 50 commits into

dev/chthonic-consecrationfrom

fix/mla-dcp-fp8-kv-cache

Commits on Jun 12, 2026

docs: document B12X vLLM binding ownership

lukealonso
and
codex
committed
165cb68
build: update CUDA dependency pins

lukealonso
and
codex
committed
ced4f6e
scripts: add local B12X serve launchers

lukealonso
and
codex
committed
6f3d496
attention: handle non-power-of-two CP LSE reductions

lukealonso
and
codex
committed
3b0f253
config: add B12X virtual TP padding

lukealonso
and
codex
committed
c609c60
compilation: prewarm B12X kernels before graph capture

lukealonso
and
codex
committed
5771a3a
config: register B12X kernel backend names

lukealonso
and
codex
committed
e87860b
kernels: add B12X FP8 linear backend

lukealonso
and
codex
committed
8171997
moe: add native B12X FP4 experts

lukealonso
and
codex
committed
5242e26
quantization: pad gated NVFP4 MoE swizzled scales

lukealonso
and
codex
committed
3b5c392
quantization: route ModelOpt mixed MXFP8 layers

lukealonso
and
codex
committed
79eede7
distributed: add optional PCIe allreduce backends

lukealonso
and
codex
committed
c1e7d46
warmup: gate autotune and prewarm DSv4 mHC kernels

lukealonso
and
codex
committed
cdbabf8
attention: add B12X sparse indexer path

lukealonso
and
codex
committed
012c1b9
attention: share sparse MLA batch metadata

lukealonso
and
codex
committed
f12c1d9
attention: add B12X sparse MLA backend

lukealonso
and
codex
committed
1336876
attention: map compressed MLA caches for DCP

lukealonso
and
codex
committed
39ec519
models: wire B12X DeepSeek V4 execution paths

lukealonso
and
codex
committed
839d492
cache: account DSV4 DCP KV memory

lukealonso
and
codex
committed
d86ba17
compilation: keep small spec decode graphs exact

lukealonso
and
codex
committed
a53bca4
spec_decode: avoid live-shape Triton specialization

lukealonso
and
codex
committed
ab949e2
spec_decode: harden GPU rejection sampler

lukealonso
and
codex
committed
6195ab8
spec_decode: support quantized DeepSeek MTP draft layers

lukealonso
and
codex
committed
92ff8cd
config: allow reasoning parsers with V2 runner

lukealonso
and
codex
committed
d9b9fbf
worker: fix KV cache memory hint

lukealonso
and
codex
committed
4265309
Tighten DeepSeek V4 MTP loading and warmup

lukealonso
and
codex
committed
2bfba29
Avoid B12X sparse indexer K copies

lukealonso
and
codex
committed
4f0fbbf
Guard draft logits writes to active requests

lukealonso
and
codex
committed
39abd3b
fix b12x modelopt nvfp4 a16 forcing

Codex
authored and
lukealonso
committed
d56d5af
Fix DCP sparse indexer decode metadata

codex
authored and
lukealonso
committed
5dc33bb
Make B12X virtual TP padding automatic

lukealonso
and
codex
committed
db15864
Use unified b12x indexer planning
lukealonso
committed
0b19b60
Assert b12x prefill uses packed indexer route
lukealonso
committed
b56cb0a
Reserve b12x MLA warmup scratch
lukealonso
committed
a0d3c9e
Support MiMo V2.5 FP4-DFlash B12X Triton runtime

codex
authored and
lukealonso
committed
d178120
Support MiMo FP4 FlashInfer Cutlass MoE

codex
authored and
lukealonso
committed
dc13c5b
Apply DFlash uniform batch prefill fix

codex
authored and
lukealonso
committed
987ef60
Port DFlash SWA support

codex
authored and
lukealonso
committed
e3c3454
Port DFlash FlashInfer backend support

codex
authored and
lukealonso
committed
945321d
Port FlashInfer metadata grouping for DFlash SWA

codex
authored and
lukealonso
committed
6e4a22f
Honor MiMo DFlash SWA config

codex
authored and
lukealonso
committed
31f1768
Route b12x MXFP4 MoE to DeepSeek-style method

authored and
lukealonso
committed
f7448bf
Fix DFlash draft fidelity to reference semantics

authored and
lukealonso
committed
c38177d
Fix DFlash long-context decode slowdown

authored and
lukealonso
committed
7703271
Allow DFlash SWA drafts beside MLA target models

authored and
lukealonso
committed
be4927a
Avoid MLA decode smem overflow on 99KB-limit GPUs

authored and
lukealonso
committed
4933b09
Port tuned TRITON_MLA decode from glm51-v6 branch

authored and
lukealonso
committed
c485569
Support DFlash on the V2 model runner

authored and
lukealonso
committed
b180f55
Add B12X paged attention backend

lukealonso
and
codex
committed
1fce5b3
attention: support quantized KV cache in MLA DCP context prefill

voipmonitor
and
claude
committed
6e0727c