* Add fused biased_grouped_topk

* add record function

* feat: automated pip install, but manual environmental variable

* fix:small bug

* feat: added requirements.txt, but encountered some other issues

* fix: run_time lib configed correctly, but _cython does not exist

* fix: worked?

* fix: no .vscode

* fix: "re-Deleted .vscode directory"

* Add fused biased_grouped_topk

* add record function

* Add fused biased_grouped_topk

* add record function

* Add fused biased_grouped_topk

* add record function

* Add fused biased_grouped_topk

* add record function

* Add fused biased_grouped_topk

* add record function

Signed-off-by: Ivan Butygin <ivan.butygin@gmail.com>

Signed-off-by: Ivan Butygin <ivan.butygin@gmail.com>

Signed-off-by: Ivan Butygin <ivan.butygin@gmail.com>

* sync models and datasets every 6 hours

Signed-off-by: mywaaagh_admin <pkwarcraft@gmail.com>

…)" (sgl-project#29)

…sed-new

[feat] enable fused qknorm and rope

…and revert necessary changes (sgl-project#20)

* refactor: enhance KL divergence method in DiagonalGaussianDistribution for flexible dimension handling and clean up DAC class by removing unused code

* refactor: update DacVAE architecture, configuration and its customized loader.

* Revert "fix: update adjust_frames parameter to False for improved multi-GPU compatibility"

* revert changes in base pipeline configs

* revert changes in configs/sample/__init__.py

* [Feature] Remove weight norm in DAC

* [Fix] Use legacy weight norm, which can be removed

* [Fix] remove weight norm at the right place

* [Chore] update test script

* Revert "[Fix] remove weight norm at the right place"

This reverts commit 3a0accbae41650e926c5828025323a12454827a4.

* Revert "[Fix] Use legacy weight norm, which can be removed"

This reverts commit eb93f20f134888adba4a5124fa1d167b93d180e7.

* Revert "[Feature] Remove weight norm in DAC"

This reverts commit aaa64abbc25112a706bf3d3604ffeac390a1d8a8.

* [Feature] Remove all weight norm from DAC modeling

---------

Co-authored-by: CloudRipple <yiyangzhang25@m.fudan.edu.cn>

* port layernorm 3d

* apply layernorm

* support for bias

* fix

* intf fix

* add support for CPU

* fix tp=3/6 padding issue in encoder vision

* fix tp=3/6 padding issue in qwen3-omni

* refactor code

* add mrope

* change attention_mask shape to use flash attn

* add kernel apply_rotary_pos_emb_cpu

* replace nn.Linear with ReplicatedLinear

* enable torch.compile

* construct mask using query.dtype instead of bool on CPU

* add fast path for sparse attention

* fix double free segfault by wrong setting of BLOCK_M

* improve extend kernel performance for long context length

* update test_extend.py

* update comment

* fix topk softmax performance issue

* port optimization for image preprocessor in Qwen2VLImageProcessorFast

* apply optimization for image preprocessor

* update docker file

* optimize conv3d used in patch embedding

* resolve conflict

* apply optimized conv3d

* apply optimization for flash_attn_varlen_func (sgl-project#19)

* port optimization for flash_attn_varlen_func

* apply flash_attn_varlen_func

* remove contiguous before rope (sgl-project#20)

* Revert "resolve conflict"

This reverts commit 7622f6d.

* fix after rebase

* Update pyproject_cpu.toml

* Update xeon.Dockerfile

* minor fix after rebase

* rope: add support for bf16 sincos (sgl-project#102)

* format

* Update xeon.Dockerfile

* odd tp for cpu

* Apply linear_gelu_linear and fix numa memory bind (sgl-project#22)

* [CPU]  Optimize small oc GEMM for Qwen3-next on CPU (sgl-project#12446)

Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>

* port linear_gelu_linear kernel

* apply linear_gelu_linear for TP=1

* fix numa memory bind

* apply parallel partition patch

---------

Co-authored-by: jianan-gu <jianan.gu@intel.com>

* Revert "Fix: test_vlm_offline_throughput output throughput (sgl-project#13279)" (sgl-project#101)

This reverts commit 7ee3e36.

* fix input dtype mismatch issue

* apply optimized layernorm

---------

Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com>
Co-authored-by: ZailiWang <zaili.wang@intel.com>
Co-authored-by: mingfeima <mingfei.ma@intel.com>
Co-authored-by: jianan-gu <jianan.gu@intel.com>

support Qwen3.5

Add Acknowledgment To the Team mates

…ocr (#20…"

This reverts commit c330b68.

Gemma 4 audio/vision tower changes their weights from `qkv_proj.weight` to `qkv_proj.linear.weight`
Gemma 4 vision pipeline have some misc changes

* converted supported-models/reward-models.mdx

* fixed reward models

* converting rerank models

…e Squeeze)

Adds two server-args flags and routes them into the Eagle3 verifier path:

  --speculative-verify-mode  rejection_sampling | typical_acceptance
  --speculative-typical-acceptance-alpha  float (default 0.8)

Defaults to "rejection_sampling" so existing deployments are unchanged.
When set to "typical_acceptance", the Eagle3 verifier in eagle_info.py
(single-layer) and eagle_info_v2.py (multi-layer) routes the alpha value
into both threshold_single and threshold_acc that the existing CUDA kernel
already consumes:

  // sgl-kernel/csrc/speculative/speculative_sampling.cuh:80
  if (coin <= prob_acc / threshold_acc || target_prob_single >= threshold_single)
    accept;

When alpha=1.0 this reduces to strict rejection sampling. When alpha<1.0
this is exactly Medusa's typical-acceptance algorithm. No CUDA changes.

Pre-flight context (Crucible umbrella):
  experiments/MiniMax-M2.5/squeeze/relaxed/B1-typical-acceptance/preflight.md

Plan reference (Crucible umbrella):
  docs/plans/minimax-squeeze.md §184-191 (Track B B1)

Squeeze pipeline coordination (Crucible umbrella):
  CLAUDE.md rules sgl-project#20/sgl-project#21/sgl-project#22

Branch policy: this commit lives on `squeeze-relaxed` only. The umbrella
crucible repo's sglang submodule pointer continues to track origin/main of
sglang. The flag will be merged to main after the alpha-sweep bench
validates per the squeeze plan's quality floor (≤3% per-dataset drift
from Exp F lossless baseline).