forked from sgl-project/sglang
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathNOTICE
More file actions
58 lines (50 loc) · 2.9 KB
/
Copy pathNOTICE
File metadata and controls
58 lines (50 loc) · 2.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
trymirai/sglang DFlash-TfM Fork
Copyright 2026 Mirai Labs.
This product is derived from SGLang (https://github.com/sgl-project/sglang),
Copyright 2023-2024 SGLang Team, licensed under the Apache License, Version 2.0.
This fork is based on SGLang commit 5169df70f6f5434026a89e33946a9f2af2044a79.
The Apache License, Version 2.0 is included in the LICENSE file.
Mirai Labs modifications relative to the upstream base include:
- DFlash-TfM / Weaver speculative decoding support, including the DFLASH_TFM
algorithm registration, server arguments, checkpoint loading, chain and tree
proposal construction, traversal verification, proposal sampling, CUDA graph
support, and target-KV materialization.
- Fast Gated Delta Network (GDN) tree verification for hybrid linear-attention
models, including block-quadratic tree verify kernels, fused Triton tree-GDN
verify kernels, tree-aware convolution, accepted-path SSM state replay/commit,
and reduced intermediate-state cache allocation.
- DFlash-family verification and scheduling extensions for proposal
probabilities, tree metadata, committed-sequence-length planning, hidden-state
capture, and DFlash-TfM serving/reproduction paths.
- Supporting runtime integration in speculative decoding, CUDA graph capture and
replay, logits/hidden-state processing, model runner plumbing, and Qwen3.5 /
Qwen3.6-family model support.
The file-level changes are recorded in git history. The main files added or
modified by this fork include:
- python/sglang/srt/speculative/dflash_tfm.py
- python/sglang/srt/speculative/dflash_info.py
- python/sglang/srt/speculative/dflash_info_v2.py
- python/sglang/srt/speculative/dflash_utils.py
- python/sglang/srt/speculative/dflash_worker_v2.py
- python/sglang/srt/speculative/spec_info.py
- python/sglang/srt/speculative/spec_registry.py
- python/sglang/srt/arg_groups/speculative_hook.py
- python/sglang/srt/server_args.py
- python/sglang/srt/layers/attention/fla/chunk_tree_verify.py
- python/sglang/srt/layers/attention/fla/gdn_tree_fused.py
- python/sglang/srt/layers/attention/fla/gdn_tree_triton.py
- python/sglang/srt/layers/attention/fla/fused_gdn_gating.py
- python/sglang/srt/layers/attention/fla/l2norm.py
- python/sglang/jit_kernel/triton/gdn_fused_proj.py
- python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
- python/sglang/srt/layers/attention/linear/gdn_backend.py
- python/sglang/srt/layers/attention/linear/utils.py
- python/sglang/srt/layers/logits_processor.py
- python/sglang/srt/mem_cache/memory_pool.py
- python/sglang/srt/model_executor/runner/base_runner.py
- python/sglang/srt/model_executor/runner/decode_cuda_graph_runner.py
- python/sglang/srt/models/qwen3_5.py
Unless otherwise stated in a file or directory, Mirai Labs modifications are
licensed under the Apache License, Version 2.0. This NOTICE file is for
attribution and change-summary purposes only and does not modify the terms of
the Apache License, Version 2.0.