[docker] Upgrade to sglang v0.5.17

docker/Dockerfile

@@ -1,12 +1,12 @@
-ARG SGLANG_IMAGE_TAG=v0.5.15.post1-cu129
+ARG SGLANG_IMAGE_TAG=v0.5.17-cu129
 FROM slimerl/sglang:${SGLANG_IMAGE_TAG} AS sglang
 
 # ======================================== Arguments =============================================
 
 ARG PATCH_VERSION=latest
 ARG MEGATRON_COMMIT=1dcf0dafa884ad52ffb243625717a3471643e087
-ARG DEEPGEMM_COMMIT=b38a77cd193cf38f670caae192310521d24343be
-ARG DEEPEP_COMMIT=6845ffd9d59126ec0030c13e0e155935a61e5b5a
+ARG DEEPGEMM_COMMIT=7ad54cbd80ebe24ad36cb5cd3729f37beed420bb
+ARG DEEPEP_COMMIT=c5c2b0cfa767b1afeb4d237c4ed7e5ff062a550d
 ARG DEEPEP_CUDA_ARCH_LIST=
 ARG FLASH_QLA_COMMIT=821fd9d37ede18fdc2a4e707fefe3770bfc32e58
 ARG TRANSFORMER_ENGINE_COMMIT=c9877beb87ad7e711e1869dd0b5062167ede447a
@@ -39,7 +39,7 @@ RUN if [ "$(python -c 'import torch; print(torch.version.cuda.split(".")[0])')"
     fi
 
 RUN pip install flash-linear-attention==0.4.2
-# FlashQLA currently requires TileLang 0.1.9, while SGLang 0.5.15.post1 on
+# FlashQLA currently requires TileLang 0.1.9, while SGLang 0.5.17 on
 # CUDA 13 requires 0.1.11.  Keep the existing CUDA 12 installation and use
 # the default FLA backend on CUDA 13 instead of downgrading SGLang's runtime.
 RUN if [ "$(python -c 'import torch; print(torch.version.cuda.split(".")[0])')" = "12" ]; then \
@@ -69,9 +69,9 @@ RUN git clone https://github.com/NVIDIA/Megatron-LM.git --recursive && \
     cd Megatron-LM && git checkout ${MEGATRON_COMMIT} && \
     pip install -e .
 
-# zhuzilin fork builds, grouped together right after Megatron-LM:
-# torch_memory_saver, plus the GLM-5 train/rollout alignment kernels
-# (DeepGEMM batch-invariant selection and the DeepEP low-latency fork).
+# Alignment builds, grouped together right after Megatron-LM:
+# torch_memory_saver, plus the GLM-5 train/rollout alignment kernels rebased
+# onto the DeepGEMM and DeepEP revisions shipped with SGLang 0.5.17.
 RUN TMS_CUDA_MAJOR="${TMS_CUDA_MAJOR:-$(python -c 'import torch; print(torch.version.cuda.split(".")[0])')}" && \
     export TMS_CUDA_MAJOR && \
     pip install git+https://github.com/zhuzilin/torch_memory_saver.git@${TMS_COMMIT} --no-cache-dir --force-reinstall
@@ -84,7 +84,8 @@ RUN git clone https://github.com/zhuzilin/DeepGEMM.git --recursive && \
     pip install --force-reinstall --no-deps dist/sgl_deep_gemm-*.whl && \
     cd /root/ && rm -rf DeepGEMM
 
-# DeepEP low-latency alignment fork.  CUDA 13 relocated the CCCL/libcu++ headers
+# DeepEP fork with low-latency FP8 alignment for SGLang 0.5.15.post1 and 0.5.17.
+# CUDA 13 relocated the CCCL/libcu++ headers
 # (<cuda/std/*>) under include/cccl, which nvshmem_tensor.h includes, so add that
 # directory to the compile search path (nvcc + host compiler) when it exists.
 RUN git clone https://github.com/zhuzilin/DeepEP.git /root/DeepEP && \

docker/README.md

@@ -5,9 +5,10 @@ We will publish 2 kinds of docker images:
 2. latest version, which aligns to `lmsysorg/sglang:latest`.
 
 current stable version is:
-- sglang v0.5.15.post1 (0b3bb0cbe31873994c9f989fddfe2f87ca839fdd), megatron dev 1dcf0dafa884ad52ffb243625717a3471643e087
+- sglang v0.5.17 (29481685462732237d80d86076d6563e1f658102), megatron dev 1dcf0dafa884ad52ffb243625717a3471643e087
 
 history versions:
+- sglang v0.5.15.post1 (0b3bb0cbe31873994c9f989fddfe2f87ca839fdd), megatron dev 1dcf0dafa884ad52ffb243625717a3471643e087
 - sglang v0.5.13 (28b095c01005d4a3a2a5b637b7d028b07fba31b2), megatron dev 1dcf0dafa884ad52ffb243625717a3471643e087
 - sglang v0.5.12.post1 (5a15cde858ea09b77116212a39356f2fc51b8584), megatron dev 1dcf0dafa884ad52ffb243625717a3471643e087
 - sglang v0.5.10.post1 (7c35342c10e201899e22fe2972d40e60da19ff3e), megatron dev 1dcf0dafa884ad52ffb243625717a3471643e087
@@ -20,39 +21,41 @@ history versions:
 The commands to build and publish:
 
 ```bash
-just release-primary   # CUDA 12 (cu129 base): publishes `latest`, `latest-cu129`, `<version>-cu129`
-just release-cu13      # CUDA 13 (cu130 base, Blackwell): publishes `latest-cu130`, `<version>-cu130`
+just release-primary   # publishes both CUDA 12 and CUDA 13 variants below
+just release-cu12      # cu129 base: `latest`, `latest-cu129`, `<version>-cu129`
+just release-cu13      # cu130 base: `latest-cu130`, `<version>-cu130`
 ```
 
 `slimerl/slime:latest` tracks the CUDA 12 build. The tag suffixes (`-cu129` /
 `-cu130`) match the SGLang base image. `docker/Dockerfile` branches on the base
-image's CUDA version; it defaults to the cu130 SGLang base, and the cu129 base
+image's CUDA version; it defaults to the cu129 SGLang base, and the cu130 base
 is selected via build args (see `docker/justfile`).
 
 To build a single image directly without publishing:
 
 ```bash
 # CUDA 12
 docker build -f docker/Dockerfile . \
-  --build-arg SGLANG_IMAGE_TAG=v0.5.15.post1-cu129 \
+  --build-arg SGLANG_IMAGE_TAG=v0.5.17-cu129 \
   -t slimerl/slime:latest-cu129
 
 # CUDA 13 (Blackwell)
 docker build -f docker/Dockerfile . \
   --build-arg DEEPEP_CUDA_ARCH_LIST='10.0 10.3' \
-  --build-arg SGLANG_IMAGE_TAG=v0.5.15.post1-cu130 \
+  --build-arg SGLANG_IMAGE_TAG=v0.5.17-cu130 \
   -t slimerl/slime:latest-cu130
 ```
 
 The following components are pinned and rebuilt in the image:
 
 - Megatron-LM `1dcf0dafa884ad52ffb243625717a3471643e087`, plus
   `docker/patch/<version>/megatron.patch`.
-- DeepGEMM `b38a77cd193cf38f670caae192310521d24343be` from the
-  `zhuzilin/DeepGEMM` batch-invariant branch, rebuilt as an SGLang-compatible wheel.
-- DeepEP `6845ffd9d59126ec0030c13e0e155935a61e5b5a` from the
-  `zhuzilin/DeepEP` `align_fp8_quantization` branch (GLM-5 low-latency alignment).
+- DeepGEMM `7ad54cbd80ebe24ad36cb5cd3729f37beed420bb` from the
+  `zhuzilin/DeepGEMM` fork. It adds batch-invariant FP8 kernels directly to
+  the 0.5.17 `v0.1.5` source baseline.
+- DeepEP `c5c2b0cfa767b1afeb4d237c4ed7e5ff062a550d` from the
+  `zhuzilin/DeepEP` fork. It merges the 0.5.17 baseline and provides aligned
+  low-latency FP8 modes for both SGLang 0.5.15.post1 and 0.5.17.
 
 For a non-default GPU architecture list, pass
 `--build-arg DEEPEP_CUDA_ARCH_LIST='<torch arch list>'`.
-

docker/justfile

@@ -1,17 +1,13 @@
-# Primary release stays on CUDA 12 (the cu129 SGLang base). Publishes the
-# unqualified `latest`, plus the rolling `latest-cu129` and the versioned
-# `<version>-cu129` tags. The tag suffix matches the SGLang base image.
-release-primary:
-    ARG_TAG_POSTFIX="-cu129" ARG_TAG_LATEST=1 ARG_BUILD_EXTRA_ARGS='--build-arg SGLANG_IMAGE_TAG=v0.5.15.post1-cu129' just _release-raw
-
-# B200/B300 (Blackwell, x86) CUDA 13 build on the cu130 SGLang base. Publishes
-# `latest-cu130` and `<version>-cu130` (tag suffix matches the SGLang base
-# image). DeepEP must be compiled for the Blackwell arch list; the Dockerfile
-# default (9.0) targets Hopper only. The semicolon separator is normalized to
-# spaces by PyTorch's TORCH_CUDA_ARCH_LIST parser and survives the unquoted
-# ARG_BUILD_EXTRA_ARGS expansion as one token.
+# Publish both supported CUDA variants; CUDA 12 owns the unqualified `latest` tag.
+release-primary: release-cu12 release-cu13
+
+# Publish CUDA 12 tags: `latest`, `latest-cu129`, and `<version>-cu129`.
+release-cu12:
+    ARG_TAG_POSTFIX="-cu129" ARG_TAG_LATEST=1 ARG_BUILD_EXTRA_ARGS='--build-arg SGLANG_IMAGE_TAG=v0.5.17-cu129' just _release-raw
+
+# Publish B200/B300 CUDA 13 tags; compile DeepEP for Blackwell via the build arg.
 release-cu13:
-    ARG_TAG_POSTFIX="-cu130" ARG_BUILD_EXTRA_ARGS="--build-arg DEEPEP_CUDA_ARCH_LIST=10.0;10.3 --build-arg SGLANG_IMAGE_TAG=v0.5.15.post1-cu130" just _release-raw
+    ARG_TAG_POSTFIX="-cu130" ARG_BUILD_EXTRA_ARGS="--build-arg DEEPEP_CUDA_ARCH_LIST=10.0;10.3 --build-arg SGLANG_IMAGE_TAG=v0.5.17-cu130" just _release-raw
 
 _release-raw:
     #!/bin/bash

docker/patch/latest/megatron.patch

@@ -735,14 +735,31 @@ diff --git a/megatron/core/transformer/moe/moe_utils.py b/megatron/core/transfor
 index 75825cd37..445b3fb84 100644
 --- a/megatron/core/transformer/moe/moe_utils.py
 +++ b/megatron/core/transformer/moe/moe_utils.py
-@@ -711,6 +711,9 @@ def topk_routing_with_score_function(
+@@ -711,6 +711,25 @@ def topk_routing_with_score_function(
                  scores, topk, num_groups, group_topk, _compute_topk
              )
 
-+    from slime.utils.routing_replay import get_routing_replay_compute_topk
++    from slime.utils.routing_replay import (
++        get_routing_replay_compute_topk,
++        maybe_sglang_fused_sigmoid_topk,
++    )
 +    compute_topk = get_routing_replay_compute_topk(compute_topk)
 +
-     if score_function == "softmax":
++    aligned_topk = maybe_sglang_fused_sigmoid_topk(
++        logits,
++        topk,
++        use_pre_softmax=use_pre_softmax,
++        num_groups=num_groups,
++        group_topk=group_topk,
++        scaling_factor=scaling_factor,
++        score_function=score_function,
++        expert_bias=expert_bias,
++        replay_active=router_replay is not None,
++    )
++    if aligned_topk is not None:
++        probs, top_indices = aligned_topk
+-    if score_function == "softmax":
++    elif score_function == "softmax":
          if use_pre_softmax:
              scores = torch.softmax(logits, dim=-1, dtype=torch.float32).type_as(logits)
 diff --git a/megatron/core/transformer/moe/router.py b/megatron/core/transformer/moe/router.py