PR #31086 model: DSV4.1 flash

model: DSV4.1 flash #31086

am17an merged 10 commits into master from aman/dsv41-no-mxfp8

model : add DeepSeek-V4.1-Flash (deepseek41)

9479ff04

deepseek41 : remove MXFP8, convert the Engram table to BF16

40414d1b

am17an requested a review 6 days ago

am17an requested a review from

pwilkin 6 days ago

am17an requested a review from

ggerganov 6 days ago

am17an requested a review from

CISC 6 days ago

am17an requested a review 6 days ago

github-actions added model

github-actions added testing

github-actions added ggml

github-actions added mtmd

github-actions added CUDA

github-actions added conversion

am17an commented on 2026-10-24

deepseek41 : compute the compressed-KV read layers once in the cache

863314bb

deepseek41 : use get_cells() for the shared-history check, drop llama…

c0280466

deepseek41 : gather the Engram rows in-graph like qwen4exp, drop the …

51e0cf68

deepseek41 : keep the Engram hash constants in hparams.ple_*, one mod…

1ea5ea2e

deepseek41 : drop the LLAMA_DSV41_KV_MIRROR toggle

65d1b65d

deepseek41 : keep il last in build_hc_mixes

ddcb043e

deepseek41 : sort the top-k rows before the sparse gather

f6435aef

tests : add deepseek41 to the Metal fusion baseline

20526926

pwilkin approved these changes on 2026-10-24

ngxson approved these changes on 2026-10-25

am17an merged 74bca395 into master 5 days ago

ggerganov commented on 2026-10-25

CISC commented on 2026-10-25

ggerganov commented on 2026-10-25

Reviewers

ngxson

pwilkin

ggerganov

CISC

Assignees

No one assigned

Labels

model testing ggml mtmd CUDA conversion

Milestone

No milestone

llama.cpp model: DSV4.1 flash #31086 Merged

model: DSV4.1 flash #31086

llama.cpp
model: DSV4.1 flash
#31086

Merged