Add DeepSeek-V4.1 vision support #48768
malaiwah
wants to merge
27 commits into
huggingface:main
from
malaiwah:add-deepseek-v41-vision
Add DeepSeek-V4.1 (deepseek_v41) text model
b410480b
Cross-engine audit fixes: QAT fake-quant, beam+engram reorder, fp8/fp…
0a2080b4
Fix remaining CI findings: TRF001 suppression placement, docs autodoc…
66e50421
DeepseekV41Config: adopt flat text-config checkpoints instead of rele…
f026c831
DeepSeek-V4.1: release-exact MXFP4 expert scales (32-channel) in fp8-…
d4078f94
DeepSeek-V4.1: review fixes — router_logits records logits, logits_to…
1872d61a
DeepSeek-V4.1: HF module naming, fused experts, mHC module, EP plan
df9f4b4a
deepseek_v41: move the engram n-gram history into the cache, bind the…
bf65a5e7
DeepSeek-V4.1: tests for the HF names (native-name load, fused fp4 ex…
8d94478f
DeepSeek-V4.1: docs for the HF naming / fused experts / fp4 load; rou…
f3624182
DeepSeek-V4.1: native-name load test also carries the ignored MTP / v…
7a03ce34
Merge branch 'wt-engram' into add-deepseek-v41
603b0422
DeepSeek-V4.1: standard device_map loading — model-level engram table…
1bd705d3
device_map: exclude no-placement params from the balanced per-device …
00b28c2c
DeepSeek-V4.1: engram table scales as weight_scale_inv so a dequantiz…
e56a35e8
DeepSeek-V4.1: engram tables as nn.Embedding sharded colwise under ex…
40bc44d1
DeepSeek-V4.1 docs: standard device_map loading, expert-parallel usag…
3b5e76f5
style: ruff format conversion_mapping.py
dd654cbe
DeepSeek-V4.1 tests: deflake beam-reorder engram test (min_new_tokens…
b3164472
fp8 quantizer: rewrite the TP/EP plans on the text config
2bff848c
DeepSeek-V4.1: gather the indexer's top-k entries instead of masking …
fdda3655
DeepSeek-V4.1 docs: measured expert-parallel numbers on 8xH200 and lo…
17df572d
DeepSeek-V4.1 docs: refresh the contribution date line (check_repo)
53144761
DeepSeek-V4.1: slow integration tests on the released checkpoint
6c073bd3
DeepSeek-V4.1: isolate uncached KV sources and correct EP limits
754bc795
DeepSeek-V4.1: fix padded compression, bare quantized loads and cache…
62d7ebd7
DeepSeek-V4.1: add vision tower, aligner and image-text-to-text support
a290b39f
Reviewers
No reviews
Assignees
No one assigned
Labels
Code agent slop
Milestone
No milestone
Login to write a write a comment.
Login via GitHub