diff --git a/docker/Dockerfile b/docker/Dockerfile
index e6884d3b5..fe7c2680a 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -11,7 +11,7 @@ ARG GRACE_BLACKWELL_DEEPEP_BRANCH=gb200_blog_part_2
 ARG HOPPER_SBO_DEEPEP_COMMIT=9f2fc4b3182a51044ae7ecb6610f7c9c3258c4d6
 ARG DEEPEP_COMMIT=9af0e0d0e74f3577af1979c9b9e1ac2cad0104ee
 ARG BUILD_AND_DOWNLOAD_PARALLEL=8
-ARG SGL_KERNEL_VERSION=0.4.2
+ARG SGL_KERNEL_VERSION=0.4.2.post1
 ARG SGL_VERSION
 ARG USE_LATEST_SGLANG=0
 ARG GDRCOPY_VERSION=2.5.1
diff --git a/python/pyproject.toml b/python/pyproject.toml
index 9da94e25f..19eab8a8b 100755
--- a/python/pyproject.toml
+++ b/python/pyproject.toml
@@ -59,7 +59,8 @@ dependencies = [
   "sentencepiece",
   "setproctitle",
   "flash-attn-4>=4.0.0b9",
-  "sglang-kernel==0.4.2",
+  "sgl-deep-gemm==0.0.1",
+  "sglang-kernel==0.4.2.post1",
   "soundfile==0.13.1",
   "tiktoken",
   "timm==1.0.16",
diff --git a/python/sglang/srt/entrypoints/engine.py b/python/sglang/srt/entrypoints/engine.py
index 0b7849a6b..1f7e79276 100644
--- a/python/sglang/srt/entrypoints/engine.py
+++ b/python/sglang/srt/entrypoints/engine.py
@@ -1170,7 +1170,7 @@ def _set_envs_and_config(server_args: ServerArgs):
         if _is_cuda:
             assert_pkg_version(
                 "sglang-kernel",
-                "0.4.2",
+                "0.4.2.post1",
                 "Please reinstall the latest version with `pip install sglang-kernel --force-reinstall`",
             )
 
diff --git a/python/sglang/srt/layers/attention/nsa/nsa_indexer.py b/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
index 87bfecc75..78dcf0aec 100644
--- a/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
+++ b/python/sglang/srt/layers/attention/nsa/nsa_indexer.py
@@ -456,10 +456,18 @@ class Indexer(MultiPlatformOp):
         # Reuse pre-computed schedule metadata if available (from init_forward_metadata),
         # otherwise fall back to computing it here.
         schedule_metadata = getattr(metadata, "paged_mqa_schedule_metadata", None)
+        # DeepGEMM release-0426 requires context_lens of shape [batch_size, next_n]
+        # to match q.shape = [batch_size, next_n, heads, head_dim]. The indexer uses
+        # next_n=1 with batch_size=N_total via q_fp8.unsqueeze(1) below, so mirror
+        # that layout here.
+        if seqlens_32.dim() == 2:
+            seqlens_32_2d = seqlens_32
+        else:
+            seqlens_32_2d = seqlens_32.unsqueeze(-1)
         if _is_cuda:
             if schedule_metadata is None:
                 schedule_metadata = deep_gemm.get_paged_mqa_logits_metadata(
-                    seqlens_32, blocksize, self.sm_count
+                    seqlens_32_2d, blocksize, self.sm_count
                 )
 
         assert len(q_fp8.shape) == 3
@@ -508,7 +516,7 @@ class Indexer(MultiPlatformOp):
                 q_fp8[:q_offset],
                 kv_cache_fp8,
                 weights[:q_offset],
-                seqlens_32,
+                seqlens_32_2d,
                 block_tables,
                 schedule_metadata,
                 max_seq_len,
diff --git a/python/sglang/srt/layers/attention/nsa_backend.py b/python/sglang/srt/layers/attention/nsa_backend.py
index f583509a7..f515228d4 100644
--- a/python/sglang/srt/layers/attention/nsa_backend.py
+++ b/python/sglang/srt/layers/attention/nsa_backend.py
@@ -67,6 +67,17 @@ else:
     )
 
 
+def _to_2d_context_lens(seqlens_32: torch.Tensor, batch_size: int) -> torch.Tensor:
+    if seqlens_32.dim() == 2:
+        return seqlens_32
+    n = seqlens_32.numel()
+    assert (
+        n % batch_size == 0
+    ), f"seqlens_32 size {n} is not a multiple of batch_size {batch_size}"
+    next_n = n // batch_size
+    return seqlens_32.view(batch_size, next_n)
+
+
 # Reuse this workspace buffer across all NSA backend instances
 global_workspace_buffer = None
 
@@ -647,8 +658,11 @@ class NativeSparseAttnBackend(
                     )
                     else cache_seqlens_int32
                 )
+                seqlens_32_2d = _to_2d_context_lens(
+                    seqlens_32, forward_batch.batch_size
+                )
                 paged_mqa_schedule_metadata = deep_gemm.get_paged_mqa_logits_metadata(
-                    seqlens_32, 64, deep_gemm.get_num_sms()
+                    seqlens_32_2d, 64, deep_gemm.get_num_sms()
                 )
             except (ImportError, ModuleNotFoundError):
                 paged_mqa_schedule_metadata = None
@@ -932,8 +946,9 @@ class NativeSparseAttnBackend(
                     )
                     else cache_seqlens_int32
                 )
+                seqlens_32_2d = _to_2d_context_lens(seqlens_32, bs)
                 paged_mqa_schedule_metadata = deep_gemm.get_paged_mqa_logits_metadata(
-                    seqlens_32, 64, deep_gemm.get_num_sms()
+                    seqlens_32_2d, 64, deep_gemm.get_num_sms()
                 )
             except (ImportError, ModuleNotFoundError):
                 paged_mqa_schedule_metadata = None
@@ -1082,8 +1097,9 @@ class NativeSparseAttnBackend(
                     )
                     else metadata.cache_seqlens_int32
                 )
+                seqlens_32_2d = _to_2d_context_lens(seqlens_32, bs)
                 new_schedule = deep_gemm.get_paged_mqa_logits_metadata(
-                    seqlens_32, 64, deep_gemm.get_num_sms()
+                    seqlens_32_2d, 64, deep_gemm.get_num_sms()
                 )
                 if metadata.paged_mqa_schedule_metadata is None:
                     metadata.paged_mqa_schedule_metadata = new_schedule
diff --git a/python/sglang/srt/layers/deep_gemm_wrapper/compile_utils.py b/python/sglang/srt/layers/deep_gemm_wrapper/compile_utils.py
index a30bf5e14..02c31cf15 100644
--- a/python/sglang/srt/layers/deep_gemm_wrapper/compile_utils.py
+++ b/python/sglang/srt/layers/deep_gemm_wrapper/compile_utils.py
@@ -196,12 +196,18 @@ def _compile_deep_gemm_one_type_all(
             kernel_type, max_m=max_m, n=n, k=k, num_groups=num_groups
         )
 
-        old_compile_mode = deep_gemm.get_compile_mode()
-        deep_gemm.set_compile_mode(1)
+        has_compile_mode_api = hasattr(deep_gemm, "get_compile_mode") and hasattr(
+            deep_gemm, "set_compile_mode"
+        )
+        if has_compile_mode_api:
+            old_compile_mode = deep_gemm.get_compile_mode()
+            deep_gemm.set_compile_mode(1)
+
         # TODO can use multi thread
         for m in tqdm(m_list, desc=f"DeepGEMM warmup"):
             executor.execute(m=m)
-        deep_gemm.set_compile_mode(old_compile_mode)
+        if has_compile_mode_api:
+            deep_gemm.set_compile_mode(old_compile_mode)
 
         # clean up input buffers
         torch.cuda.current_stream().synchronize()
@@ -300,7 +306,7 @@ class _GroupedContWarmupExecutor(_BaseWarmupExecutor):
             (self.lhs_q[:m], self.lhs_s[:m]),
             (self.rhs_q, self.rhs_s),
             self.out[:m],
-            m_indices=self.m_indices[:m],
+            self.m_indices[:m],
         )
 
 
diff --git a/python/sglang/srt/layers/deep_gemm_wrapper/configurer.py b/python/sglang/srt/layers/deep_gemm_wrapper/configurer.py
index 6ae5ee431..de433fe2d 100644
--- a/python/sglang/srt/layers/deep_gemm_wrapper/configurer.py
+++ b/python/sglang/srt/layers/deep_gemm_wrapper/configurer.py
@@ -18,6 +18,8 @@ def _compute_enable_deep_gemm():
     sm_version = get_device_sm()
     if (_is_cuda and sm_version < 90) or (_is_musa and sm_version < 31):
         return False
+    if not (_is_cuda or _is_musa):
+        return False
 
     try:
         import deep_gemm  # noqa: F401
diff --git a/python/sglang/srt/layers/quantization/fp8_utils.py b/python/sglang/srt/layers/quantization/fp8_utils.py
index b39da5267..3bb83c111 100755
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -1286,6 +1286,19 @@ def transform_scale_ue8m0(sf, mn, use_torch_impl: bool = False):
 
     sf = sf.index_select(-2, torch.arange(mn, device=sf.device) // 128)
     sf = get_mn_major_tma_aligned_packed_ue8m0_tensor(sf)
+
+    # In sgl-deep-gemm, the C++ deepgemm path returns through DLPack which collapses the stride
+    # of size-1 trailing dims to 1 (happens when packed_sf_k == 1, i.e.
+    # K <= block_k * 4). Restore the TMA-aligned stride so the deepgemm
+    # assertion sf.stride(-1) == get_tma_aligned_size(mn, element_size) holds.
+    if not use_torch_impl and sf.shape[-1] == 1:
+        from deep_gemm.utils import get_tma_aligned_size
+
+        aligned_mn = get_tma_aligned_size(sf.shape[-2], sf.element_size())
+        if sf.stride(-1) != aligned_mn:
+            new_stride = list(sf.stride())
+            new_stride[-1] = aligned_mn
+            sf = sf.as_strided(sf.shape, tuple(new_stride))
     return sf
 
 
diff --git a/scripts/ci/cuda/ci_install_dependency.sh b/scripts/ci/cuda/ci_install_dependency.sh
index 66a37310d..5d8e0cd69 100755
--- a/scripts/ci/cuda/ci_install_dependency.sh
+++ b/scripts/ci/cuda/ci_install_dependency.sh
@@ -294,6 +294,12 @@ install_sglang_kernel() {
     else
         echo "CUSTOM_BUILD_SGL_KERNEL=true: keeping freshly built sgl-kernel wheel."
     fi
+    SGL_DEEP_GEMM_VERSION=$(grep -Po -m1 '(?<=sgl-deep-gemm==)[0-9A-Za-z\.\-]+' python/pyproject.toml)
+    if [ "$CU_MAJOR" = "13" ]; then
+        $PIP_CMD install "sgl-deep-gemm==${SGL_DEEP_GEMM_VERSION}" --force-reinstall $PIP_INSTALL_SUFFIX
+    else
+        $PIP_CMD install "https://github.com/sgl-project/whl/releases/download/v${SGL_DEEP_GEMM_VERSION}/sgl_deep_gemm-${SGL_DEEP_GEMM_VERSION}+cu129-py3-none-manylinux2014_$(uname -m).whl" --force-reinstall $PIP_INSTALL_SUFFIX
+    fi
 
     mark_step_done "${FUNCNAME[0]}"
 }
diff --git a/scripts/ci/cuda/warmup_deep_gemm.py b/scripts/ci/cuda/warmup_deep_gemm.py
index b1844bc59..0e8a0442d 100644
--- a/scripts/ci/cuda/warmup_deep_gemm.py
+++ b/scripts/ci/cuda/warmup_deep_gemm.py
@@ -235,8 +235,11 @@ def compile_one_shape(kernel_type, n, k, num_groups, m_list):
         )
         m_list = [m for m in m_list if m <= max_m]
 
-    old_mode = deep_gemm.get_compile_mode()
-    deep_gemm.set_compile_mode(1)
+    get_compile_mode = getattr(deep_gemm, "get_compile_mode", None)
+    set_compile_mode = getattr(deep_gemm, "set_compile_mode", None)
+    old_mode = get_compile_mode() if get_compile_mode is not None else None
+    if set_compile_mode is not None:
+        set_compile_mode(1)
     try:
         if kernel_type == "NORMAL":
             lhs_q, lhs_s = _empty_token_fp8((max_m, k))
@@ -255,7 +258,7 @@ def compile_one_shape(kernel_type, n, k, num_groups, m_list):
                     (lhs_q[:m], lhs_s[:m]),
                     (rhs_q, rhs_s),
                     out[:m],
-                    m_indices=m_indices[:m],
+                    m_indices[:m],
                 )
 
         elif kernel_type == "MASKED":
@@ -274,7 +277,8 @@ def compile_one_shape(kernel_type, n, k, num_groups, m_list):
                     expected_m=m,
                 )
     finally:
-        deep_gemm.set_compile_mode(old_mode)
+        if set_compile_mode is not None and old_mode is not None:
+            set_compile_mode(old_mode)
 
     torch.cuda.current_stream().synchronize()
     torch.cuda.empty_cache()
diff --git a/test/registered/8-gpu-models/test_dsa_models_mtp.py b/test/registered/8-gpu-models/test_dsa_models_mtp.py
index 05cd47632..9bb05f36c 100644
--- a/test/registered/8-gpu-models/test_dsa_models_mtp.py
+++ b/test/registered/8-gpu-models/test_dsa_models_mtp.py
@@ -16,7 +16,11 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=1048, suite="stage-c-test-8-gpu-h200")
+register_cuda_ci(
+    est_time=1048,
+    suite="stage-c-test-8-gpu-h200",
+    disabled="Disabled due to #24268. Should be fixed soon.",
+)
 
 FULL_DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 GLM5_MODEL_PATH = "zai-org/GLM-5-FP8"
diff --git a/test/registered/cp/test_deepseek_v32_cp_single_node.py b/test/registered/cp/test_deepseek_v32_cp_single_node.py
index 55fa21907..dcfcb4d78 100644
--- a/test/registered/cp/test_deepseek_v32_cp_single_node.py
+++ b/test/registered/cp/test_deepseek_v32_cp_single_node.py
@@ -13,7 +13,11 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=616, suite="stage-c-test-deepep-8-gpu-h200")
+register_cuda_ci(
+    est_time=616,
+    suite="stage-c-test-deepep-8-gpu-h200",
+    disabled="Disabled due to #24268. Should be fixed soon.",
+)
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
 
diff --git a/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py b/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py
index 1cf38cdf7..14793c17a 100644
--- a/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py
+++ b/test/registered/quant/test_deepseek_v32_fp4_mtp_4gpu.py
@@ -15,7 +15,11 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=1060, suite="stage-c-test-4-gpu-b200")
+register_cuda_ci(
+    est_time=1060,
+    suite="stage-c-test-4-gpu-b200",
+    disabled="Disabled due to #24268. Should be fixed soon.",
+)
 
 FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3.2-NVFP4"
 SERVER_LAUNCH_TIMEOUT = 1200