Support Hy4-preview (#36805)

Co-authored-by: BBuf <1182563586@qq.com>
Co-authored-by: alphabetc1 <2508695655@qq.com>

python/sglang/kernels/ops/layernorm/__init__.py

@@ -514,6 +514,8 @@ _PHASE25_KERNELS = [
     ("gemma4_fused_ops", "gemma4_fused_routing", "triton"),
     ("gemma4_fused_ops", "gemma_qkv_rmsnorm", "triton"),
     ("mhc_head", "fused_hc_head", "triton"),
+    ("hy4_ihc", "fused_hy4_ihc_pre", "triton"),
+    ("hy4_ihc", "fused_hy4_ihc_post", "triton"),
 ]
 for _mod, _fn, _bk in _PHASE25_KERNELS:
     register_kernel(

python/sglang/kernels/ops/layernorm/hy4_ihc.py

@@ -0,0 +1,443 @@
+from __future__ import annotations
+
+import functools
+import logging
+
+import torch
+import triton
+import triton.language as tl
+
+logger = logging.getLogger(__name__)
+
+
+_HPC_IHC_CAPABILITIES = ((9, 0), (10, 0), (10, 3))
+_HPC_IHC_HC_MULTS = (4,)
+_HPC_IHC_HIDDEN_SIZES = (4096, 6144)
+
+
+@functools.lru_cache(maxsize=None)
+def _hpc_ihc_op(op_name: str, hc_mult: int, hidden_size: int):
+    try:
+        import hpc
+    except ImportError:
+        return None
+
+    op = getattr(hpc, op_name, None)
+    if op is None:
+        logger.info(
+            "HY4 iHC: the installed hpc build (%s) has no %s; using the "
+            "in-tree Triton kernels.",
+            getattr(hpc, "__version__", "unknown"),
+            op_name,
+        )
+        return None
+
+    from sglang.srt.utils import get_device_capability
+
+    cap = get_device_capability()
+    if cap not in _HPC_IHC_CAPABILITIES:
+        logger.warning(
+            "HY4 iHC: hpc.%s is unavailable on sm%s%s.",
+            op_name,
+            *cap,
+        )
+        return None
+    if hc_mult not in _HPC_IHC_HC_MULTS or hidden_size not in _HPC_IHC_HIDDEN_SIZES:
+        logger.warning(
+            "HY4 iHC: hpc.%s is instantiated for hc_mult in %s and hidden_size "
+            "in %s, got %d / %d.",
+            op_name,
+            _HPC_IHC_HC_MULTS,
+            _HPC_IHC_HIDDEN_SIZES,
+            hc_mult,
+            hidden_size,
+        )
+        return None
+
+    logger.info("HY4 iHC: using hpc.%s.", op_name)
+    return op
+
+
+@triton.jit
+def _hy4_ihc_pre_stage1(
+    x_ptr,
+    fn_ptr,
+    part_ptr,
+    K_TOTAL: tl.constexpr,
+    HC_MULT: tl.constexpr,
+    HC_POW2: tl.constexpr,
+    NSPLIT: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    PART_STRIDE: tl.constexpr,
+):
+    pid_t = tl.program_id(0).to(tl.int64)
+    pid_s = tl.program_id(1)
+    m_idx = tl.arange(0, HC_POW2)
+    m_mask = m_idx < HC_MULT
+
+    k_offs = pid_s * BLOCK_K + tl.arange(0, BLOCK_K)
+    k_mask = k_offs < K_TOTAL
+
+    x_tile = tl.load(x_ptr + pid_t * K_TOTAL + k_offs, mask=k_mask, other=0.0).to(
+        tl.float32
+    )
+    sumsq = tl.sum(x_tile * x_tile, axis=0)
+
+    fn_offs = m_idx[:, None] * K_TOTAL + k_offs[None, :]
+    fn_mask = m_mask[:, None] & k_mask[None, :]
+    mix_pre = tl.sum(
+        tl.load(fn_ptr + fn_offs, mask=fn_mask, other=0.0) * x_tile[None, :], axis=1
+    )
+    mix_post = tl.sum(
+        tl.load(fn_ptr + HC_MULT * K_TOTAL + fn_offs, mask=fn_mask, other=0.0)
+        * x_tile[None, :],
+        axis=1,
+    )
+
+    base = part_ptr + (pid_t * NSPLIT + pid_s) * PART_STRIDE
+    tl.store(base, sumsq)
+    tl.store(base + 1 + m_idx, mix_pre, mask=m_mask)
+    tl.store(base + 1 + HC_POW2 + m_idx, mix_post, mask=m_mask)
+
+
+@triton.jit
+def _hy4_ihc_pre_stage2(
+    x_ptr,
+    part_ptr,
+    scale_ptr,
+    base_ptr,
+    y_ptr,
+    post_ptr,
+    hidden_size: tl.constexpr,
+    K_TOTAL: tl.constexpr,
+    HC_MULT: tl.constexpr,
+    HC_POW2: tl.constexpr,
+    NSPLIT: tl.constexpr,
+    PART_STRIDE: tl.constexpr,
+    BLOCK_D: tl.constexpr,
+    magnitude: tl.constexpr,
+    norm_eps: tl.constexpr,
+    hc_eps: tl.constexpr,
+):
+    pid_t = tl.program_id(0).to(tl.int64)
+    pid_d = tl.program_id(1)
+    m_idx = tl.arange(0, HC_POW2)
+    m_mask = m_idx < HC_MULT
+
+    # The single-CTA kernel folded the BLOCK_K tiles into one accumulator with
+    # '+='; replay that ascending order over the partials so the fp32 sum is
+    # bit-identical.
+    row = part_ptr + pid_t * NSPLIT * PART_STRIDE
+    sumsq = tl.zeros((), dtype=tl.float32)
+    mix_pre = tl.zeros((HC_POW2,), dtype=tl.float32)
+    mix_post = tl.zeros((HC_POW2,), dtype=tl.float32)
+    for s in tl.static_range(NSPLIT):
+        b = row + s * PART_STRIDE
+        sumsq += tl.load(b)
+        mix_pre += tl.load(b + 1 + m_idx, mask=m_mask, other=0.0)
+        mix_post += tl.load(b + 1 + HC_POW2 + m_idx, mask=m_mask, other=0.0)
+
+    rsqrt = tl.rsqrt(sumsq / K_TOTAL + norm_eps)
+    scale_pre = tl.load(scale_ptr)
+    scale_post = tl.load(scale_ptr + 1)
+    base_pre = tl.load(base_ptr + m_idx, mask=m_mask, other=0.0)
+    base_post = tl.load(base_ptr + HC_MULT + m_idx, mask=m_mask, other=0.0)
+
+    pre = tl.sigmoid(mix_pre * rsqrt * scale_pre + base_pre) + hc_eps
+    if pid_d == 0:
+        post = (
+            magnitude * tl.sigmoid(mix_post * rsqrt * scale_post + base_post) + hc_eps
+        )
+        tl.store(post_ptr + pid_t * HC_MULT + m_idx, post, mask=m_mask)
+
+    x_row = x_ptr + pid_t * K_TOTAL
+    d_offs = pid_d * BLOCK_D + tl.arange(0, BLOCK_D)
+    d_mask = d_offs < hidden_size
+    y_block = tl.zeros((BLOCK_D,), dtype=tl.float32)
+    for m in tl.static_range(HC_MULT):
+        x_m = tl.load(x_row + m * hidden_size + d_offs, mask=d_mask, other=0.0)
+        pre_m = tl.sum(tl.where(m_idx == m, pre, 0.0), axis=0)
+        y_block += pre_m * x_m.to(tl.float32)
+    tl.store(
+        y_ptr + pid_t * hidden_size + d_offs,
+        y_block.to(y_ptr.dtype.element_ty),
+        mask=d_mask,
+    )
+
+
+@triton.jit
+def _hy4_ihc_post_kernel(
+    out_ptr,
+    res_ptr,
+    post_ptr,
+    y_ptr,
+    hidden_size: tl.constexpr,
+    HC_MULT: tl.constexpr,
+    HC_POW2: tl.constexpr,
+    BLOCK_D: tl.constexpr,
+):
+    pid_t = tl.program_id(0).to(tl.int64)
+    pid_d = tl.program_id(1)
+
+    m_idx = tl.arange(0, HC_POW2)
+    m_mask = m_idx < HC_MULT
+    post = tl.load(post_ptr + pid_t * HC_MULT + m_idx, mask=m_mask, other=0.0)
+
+    d_offs = pid_d * BLOCK_D + tl.arange(0, BLOCK_D)
+    d_mask = d_offs < hidden_size
+    out_block = tl.load(out_ptr + pid_t * hidden_size + d_offs, mask=d_mask, other=0.0)
+    out_block = out_block.to(tl.float32)
+
+    res_row = res_ptr + pid_t * HC_MULT * hidden_size
+    y_row = y_ptr + pid_t * HC_MULT * hidden_size
+    for m in tl.static_range(HC_MULT):
+        res_block = tl.load(res_row + m * hidden_size + d_offs, mask=d_mask, other=0.0)
+        post_m = tl.sum(tl.where(m_idx == m, post, 0.0), axis=0)
+        y_block = post_m * out_block + res_block.to(tl.float32)
+        tl.store(
+            y_row + m * hidden_size + d_offs,
+            y_block.to(y_ptr.dtype.element_ty),
+            mask=d_mask,
+        )
+
+
+def fused_hy4_ihc_pre(
+    x: torch.Tensor,
+    hc_fn: torch.Tensor,
+    hc_scale: torch.Tensor,
+    hc_base: torch.Tensor,
+    magnitude: float,
+    norm_eps: float,
+    hc_eps: float,
+    rms_weight: torch.Tensor | None = None,
+    rms_eps: float = 0.0,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    assert x.dim() == 3, f"x must be 3D (T, hc_mult, hidden_size), got {x.shape}"
+    assert hc_fn.dtype == torch.float32
+    assert hc_scale.dtype == torch.float32 and hc_base.dtype == torch.float32
+
+    x = x.contiguous()
+    hc_fn = hc_fn.contiguous()
+    T, hc_mult, hidden_size = x.shape
+    k_total = hc_mult * hidden_size
+    assert hc_fn.shape == (2 * hc_mult, k_total)
+    assert hc_base.shape == (2 * hc_mult,)
+    assert hc_scale.shape == (2,)
+
+    if T == 0:
+        return (
+            torch.empty((0, hidden_size), dtype=x.dtype, device=x.device),
+            torch.empty((0, hc_mult), dtype=torch.float32, device=x.device),
+        )
+
+    hpc_op = _hpc_ihc_op("fuse_ihc_pre", hc_mult, hidden_size)
+    if hpc_op is not None:
+        return hpc_op(
+            x,
+            hc_fn,
+            hc_scale,
+            hc_base,
+            norm_eps,
+            hc_eps,
+            magnitude,
+            rms_weight,
+            rms_eps,
+            True,
+        )
+
+    y = torch.empty((T, hidden_size), dtype=x.dtype, device=x.device)
+    post = torch.empty((T, hc_mult), dtype=torch.float32, device=x.device)
+    BLOCK_K = 1024
+    BLOCK_D = 1024
+    hc_pow2 = triton.next_power_of_2(hc_mult)
+    # One BLOCK_K tile per CTA: the partials then replay the old kernel's
+    # per-tile accumulation order exactly.
+    nsplit = triton.cdiv(k_total, BLOCK_K)
+    part_stride = 1 + 2 * hc_pow2
+    part = torch.empty((T, nsplit, part_stride), dtype=torch.float32, device=x.device)
+
+    _hy4_ihc_pre_stage1[(T, nsplit)](
+        x,
+        hc_fn,
+        part,
+        K_TOTAL=k_total,
+        HC_MULT=hc_mult,
+        HC_POW2=hc_pow2,
+        NSPLIT=nsplit,
+        BLOCK_K=BLOCK_K,
+        PART_STRIDE=part_stride,
+        num_warps=8,
+        # Disable FMA: eager rounds the fp32 product before summation.
+        enable_fp_fusion=False,
+    )
+    _hy4_ihc_pre_stage2[(T, triton.cdiv(hidden_size, BLOCK_D))](
+        x,
+        part,
+        hc_scale,
+        hc_base,
+        y,
+        post,
+        hidden_size=hidden_size,
+        K_TOTAL=k_total,
+        HC_MULT=hc_mult,
+        HC_POW2=hc_pow2,
+        NSPLIT=nsplit,
+        PART_STRIDE=part_stride,
+        BLOCK_D=BLOCK_D,
+        magnitude=magnitude,
+        norm_eps=norm_eps,
+        hc_eps=hc_eps,
+        num_warps=4,
+        enable_fp_fusion=False,
+    )
+    if rms_weight is not None:
+        y_float = y.float()
+        y = (
+            y_float
+            * torch.rsqrt(y_float.square().mean(dim=-1, keepdim=True) + rms_eps)
+            * rms_weight.float()
+        ).to(y.dtype)
+    return y, post
+
+
+def fused_hy4_ihc_post(
+    output: torch.Tensor,
+    residual: torch.Tensor,
+    post: torch.Tensor,
+) -> torch.Tensor:
+    assert output.dim() == 2, f"output must be 2D (T, hidden_size), got {output.shape}"
+    assert post.dtype == torch.float32
+
+    output = output.contiguous()
+    residual = residual.contiguous()
+    T, hidden_size = output.shape
+    hc_mult = post.shape[-1]
+    assert residual.shape == (T, hc_mult, hidden_size)
+    assert post.shape == (T, hc_mult)
+
+    if T == 0:
+        return torch.empty(
+            (0, hc_mult, hidden_size), dtype=output.dtype, device=output.device
+        )
+
+    hpc_op = _hpc_ihc_op("fuse_ihc_post", hc_mult, hidden_size)
+    if hpc_op is not None:
+        return hpc_op(output, residual, post)
+
+    y = torch.empty((T, hc_mult, hidden_size), dtype=output.dtype, device=output.device)
+    BLOCK_D = 1024
+    grid = (T, triton.cdiv(hidden_size, BLOCK_D))
+    _hy4_ihc_post_kernel[grid](
+        output,
+        residual,
+        post.contiguous(),
+        y,
+        hidden_size=hidden_size,
+        HC_MULT=hc_mult,
+        HC_POW2=triton.next_power_of_2(hc_mult),
+        BLOCK_D=BLOCK_D,
+        num_warps=4,
+        enable_fp_fusion=False,
+    )
+    return y
+
+
+def fused_hy4_ihc_post_pre(
+    output: torch.Tensor,
+    residual: torch.Tensor,
+    post: torch.Tensor,
+    hc_fn: torch.Tensor,
+    hc_scale: torch.Tensor,
+    hc_base: torch.Tensor,
+    magnitude: float,
+    norm_eps: float,
+    hc_eps: float,
+    rms_weight: torch.Tensor | None = None,
+    rms_eps: float = 0.0,
+) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    num_tokens, hidden_size = output.shape
+    hc_mult = residual.shape[1]
+    if num_tokens == 0:
+        return (
+            torch.empty_like(residual),
+            torch.empty_like(output),
+            torch.empty((0, hc_mult), dtype=torch.float32, device=output.device),
+        )
+
+    hpc_op = _hpc_ihc_op("fuse_ihc_post_pre", hc_mult, hidden_size)
+    if hpc_op is not None:
+        return hpc_op(
+            output.contiguous(),
+            residual.contiguous(),
+            post.contiguous(),
+            hc_fn.contiguous(),
+            hc_scale.contiguous(),
+            hc_base.contiguous(),
+            norm_eps,
+            hc_eps,
+            magnitude,
+            rms_weight,
+            rms_eps,
+            True,
+        )
+
+    next_residual = fused_hy4_ihc_post(output, residual, post)
+    reduced, next_post = fused_hy4_ihc_pre(
+        next_residual,
+        hc_fn,
+        hc_scale,
+        hc_base,
+        magnitude,
+        norm_eps,
+        hc_eps,
+        rms_weight,
+        rms_eps,
+    )
+    return next_residual, reduced, next_post
+
+
+def fused_hy4_ihc_head(
+    hidden_states: torch.Tensor,
+    hc_fn: torch.Tensor,
+    hc_scale: torch.Tensor,
+    hc_base: torch.Tensor,
+    norm_eps: float,
+    hc_eps: float,
+    rms_weight: torch.Tensor | None = None,
+    rms_eps: float = 0.0,
+) -> torch.Tensor:
+    num_tokens, hc_mult, hidden_size = hidden_states.shape
+    if num_tokens == 0:
+        return torch.empty(
+            (0, hidden_size), dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+    hpc_op = _hpc_ihc_op("fuse_ihc_head", hc_mult, hidden_size)
+    if hpc_op is not None:
+        return hpc_op(
+            hidden_states.contiguous(),
+            hc_fn.contiguous(),
+            hc_scale.contiguous(),
+            hc_base.contiguous(),
+            norm_eps,
+            hc_eps,
+            rms_weight,
+            rms_eps,
+            True,
+        )
+
+    flat = hidden_states.flatten(1).float()
+    scale = torch.rsqrt(flat.square().mean(-1, keepdim=True) + norm_eps)
+    gates = torch.nn.functional.linear(flat, hc_fn) * scale
+    gates = torch.sigmoid(gates * hc_scale + hc_base) + hc_eps
+    output = torch.sum(gates.unsqueeze(-1) * hidden_states.float(), dim=1).to(
+        hidden_states.dtype
+    )
+    if rms_weight is not None:
+        output_float = output.float()
+        output = (
+            output_float
+            * torch.rsqrt(output_float.square().mean(dim=-1, keepdim=True) + rms_eps)
+            * rms_weight.float()
+        ).to(output.dtype)
+    return output

python/sglang/kernels/ops/moe/ep_moe_kernels.py

@@ -1155,9 +1155,12 @@ def ep_scatter(
     output_index: torch.Tensor,
     scale_ue8m0: bool = False,
     quant_block_size: int = 128,
+    expert_alignment: int = 128,
     expert_start: int = 0,
 ):
-    BLOCK_E = 128  # token num of per expert is aligned to 128
+    # tl.arange needs pow2, and the kernel's unmasked stores need BLOCK_E to
+    # divide the expert_alignment-padded segments; lowbit satisfies both.
+    BLOCK_E = expert_alignment & -expert_alignment
     BLOCK_D = quant_block_size  # block size of quantization
     num_warps = 8
     num_experts = num_recv_tokens_per_expert.shape[0]

python/sglang/kernels/ops/moe/triton_pad_expert_counts.py

@@ -0,0 +1,41 @@
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _pad_expert_counts_kernel(
+    counts_ptr,
+    out_ptr,
+    num_experts,
+    all_tokens,
+    BLOCK_E: tl.constexpr,
+    NE_POW2: tl.constexpr,
+):
+    i = tl.arange(0, NE_POW2)
+    m = i < num_experts
+    c = tl.load(counts_ptr + i, mask=m, other=0).to(tl.int32)
+    padded = ((c + BLOCK_E - 1) // BLOCK_E) * BLOCK_E
+    padded = tl.where(m, padded, 0)
+    # The trailing segment absorbs the difference so the total stays
+    # graph-static; its m_indices remain -1 and DeepGEMM skips those rows.
+    slack = all_tokens - tl.sum(padded, axis=0)
+    padded = tl.where(i == num_experts - 1, padded + slack, padded)
+    tl.store(out_ptr + i, padded, mask=m)
+
+
+def pad_expert_counts(
+    counts: torch.Tensor, block_e: int, all_tokens: int
+) -> torch.Tensor:
+    ne = counts.numel()
+    out = torch.empty(ne, dtype=torch.int32, device=counts.device)
+    _pad_expert_counts_kernel[(1,)](
+        counts,
+        out,
+        ne,
+        all_tokens,
+        BLOCK_E=block_e,
+        NE_POW2=triton.next_power_of_2(ne),
+        num_warps=4,
+    )
+    return out

python/sglang/kernels/ops/moe/triton_sigmoid_gate_mul.py

@@ -37,6 +37,8 @@ def _sigmoid_gate_mul_kernel(
 
 def sigmoid_gate_mul(x: torch.Tensor, gate: torch.Tensor) -> torch.Tensor:
     """Compute ``x * sigmoid(gate)`` in a single fused kernel (same-shape)."""
+    assert x.shape == gate.shape, f"shape mismatch: {x.shape=} {gate.shape=}"
+    assert x.is_contiguous() and gate.is_contiguous(), "inputs must be contiguous"
     out = torch.empty_like(x)
     n = x.numel()
     grid = lambda meta: (triton.cdiv(n, meta["BLOCK_SIZE"]),)

python/sglang/srt/arg_groups/model_hook.py

@@ -44,6 +44,30 @@ from sglang.srt.utils.common import (
 logger = logging.getLogger(__name__)
 
 
+def _validate_dsa_tbo_index_sharing(server_args: Any, hf_config: Any) -> None:
+    cfg = resolving_view(server_args)
+    if not cfg.enable_two_batch_overlap:
+        return
+
+    index_topk_freq = getattr(hf_config, "index_topk_freq", 1) or 1
+    index_topk_pattern = getattr(hf_config, "index_topk_pattern", None)
+    indexer_types = getattr(hf_config, "indexer_types", None)
+    if (
+        index_topk_freq > 1
+        or (index_topk_pattern is not None and "S" in index_topk_pattern)
+        or (indexer_types is not None and "shared" in indexer_types)
+    ):
+        raise ValueError(
+            "--enable-two-batch-overlap is not supported with DSA "
+            "index-topk sharing: the TBO op path does not propagate topk "
+            "indices across layers, so shared layers would run sparse "
+            "attention without indices. Got "
+            f"index_topk_freq={index_topk_freq!r}, "
+            f"index_topk_pattern={index_topk_pattern!r}, and "
+            f"indexer_types={indexer_types!r}."
+        )
+
+
 def _rocm_fp8_wo_a_supported() -> bool:
     """True when ROCm can run the DeepSeek-V4 fp8 wo_a GEMM (gfx950 + aiter)."""
     try:
@@ -159,6 +183,8 @@ def handle_model_specific_adjustments(server_args: Any):
         "MistralLarge3ForCausalLM",
         "PixtralForConditionalGeneration",
         "GlmMoeDsaForCausalLM",
+        "HYV4ForCausalLM",
+        "HYV4ForCausalLMNextN",
         "LongcatFlashForCausalLM",
         "Dots3NoteForCausalLM",
     ]:
@@ -181,19 +207,7 @@ def handle_model_specific_adjustments(server_args: Any):
             # The "dsa" attention fill moved to the override registry
             # (arg_groups/overrides.py: _deepseek_family_overrides).
 
-            index_topk_freq = getattr(hf_config, "index_topk_freq", 1) or 1
-            index_topk_pattern = getattr(hf_config, "index_topk_pattern", None)
-            if cfg.enable_two_batch_overlap and (
-                index_topk_freq > 1
-                or (index_topk_pattern is not None and "S" in index_topk_pattern)
-            ):
-                raise ValueError(
-                    "--enable-two-batch-overlap is not supported with DSA "
-                    "index-topk sharing (index_topk_freq > 1 or an "
-                    "index_topk_pattern containing shared layers): the TBO op "
-                    "path does not propagate topk indices across layers, so "
-                    "shared layers would run sparse attention without indices."
-                )
+            _validate_dsa_tbo_index_sharing(server_args, hf_config)
 
             if (
                 not get_platform().is_npu and not get_platform().is_xpu

python/sglang/srt/arg_groups/model_overrides/deepseek_v2.py

@@ -1,6 +1,6 @@
 """Config-time override declarations for deepseek_v2.
 
-Architectures: DeepseekV32ForCausalLM, DeepseekV3ForCausalLM, Dots3NoteForCausalLM, GlmMoeDsaForCausalLM, KimiK25ForConditionalGeneration, LongcatFlashForCausalLM, LongcatFlashForCausalLMNextN, MistralLarge3ForCausalLM, PixtralForConditionalGeneration.
+Architectures: DeepseekV32ForCausalLM, DeepseekV3ForCausalLM, Dots3NoteForCausalLM, GlmMoeDsaForCausalLM, HYV4ForCausalLM, HYV4ForCausalLMNextN, KimiK25ForConditionalGeneration, LongcatFlashForCausalLM, LongcatFlashForCausalLMNextN, MistralLarge3ForCausalLM, PixtralForConditionalGeneration.
 """
 
 import logging
@@ -24,6 +24,8 @@ logger = logging.getLogger(__name__)
     "MistralLarge3ForCausalLM",
     "PixtralForConditionalGeneration",
     "GlmMoeDsaForCausalLM",
+    "HYV4ForCausalLM",
+    "HYV4ForCausalLMNextN",
     "LongcatFlashForCausalLM",
     "LongcatFlashForCausalLMNextN",
     "Dots3NoteForCausalLM",
@@ -35,10 +37,50 @@ def _deepseek_family_overrides(server_args: Any, hf_config: Any) -> dict:
     before it by _set_default_dsa_kv_cache_dtype) and the env writes stay in
     the branch."""
     cfg = resolving_view(server_args)
-    from sglang.srt.configs.model_config import is_deepseek_dsa
+    from sglang.srt.configs.model_config import (
+        is_deepseek_dsa,
+        unwrap_modelopt_quantization_config,
+    )
+
+    model_arch = (getattr(hf_config, "architectures", None) or [None])[0]
+    if model_arch in ("HYV4ForCausalLM", "HYV4ForCausalLMNextN"):
+        if cfg.enable_prefill_cp:
+            raise ValueError(
+                "--enable-prefill-cp is not supported for HYV4 because its "
+                "attention path does not implement DSA context-parallel metadata "
+                f"and sharding. Got architecture={model_arch!r} and "
+                f"enable_prefill_cp={cfg.enable_prefill_cp!r}."
+            )
+        dcp_size = getattr(cfg, "dcp_size", 1)
+        if dcp_size > 1:
+            raise ValueError(
+                "--dcp-size > 1 is not supported for HYV4 because decode context "
+                "parallelism gathers query heads across DCP ranks but does not "
+                "provide single-owner semantics for learnable attention sinks. "
+                f"Got architecture={model_arch!r} and dcp_size={dcp_size!r}."
+            )
 
     overrides: Dict[str, Any] = {}
 
+    if model_arch in ("HYV4ForCausalLM", "HYV4ForCausalLMNextN"):
+        quant_cfg = getattr(hf_config, "quantization_config", None) or {}
+        quant_algo = unwrap_modelopt_quantization_config(quant_cfg).get(
+            "quant_algo", ""
+        )
+        if str(quant_algo).upper() == "MXFP8":
+            from sglang.srt.layers import deep_gemm_wrapper
+
+            # auto would otherwise select an unqualified FP8/MoE path for HYV4 MXFP8.
+            if deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM:
+                if cfg.moe_runner_backend == "auto":
+                    overrides["moe_runner_backend"] = "deep_gemm"
+                if cfg.fp8_gemm_runner_backend == "auto":
+                    overrides["fp8_gemm_runner_backend"] = "deep_gemm"
+                if overrides:
+                    logger.info(
+                        "HYV4 MXFP8: defaulting MoE/FP8 GEMM backends to deep_gemm."
+                    )
+
     if is_deepseek_dsa(hf_config):  # DeepSeek 3.2/GLM 5
         # Set attention backend for DeepSeek
         if is_attention_backend_not_set(cfg):

python/sglang/srt/arg_groups/overrides.py

@@ -628,8 +628,16 @@ def _dsa_kv_cache_dtype_default(view: Any) -> dict:
         )
 
     kv_cache_dtype = view.kv_cache_dtype
+    has_attention_sinks = bool(getattr(hf_config, "learnable_sink", False))
+    if has_attention_sinks and kv_cache_dtype not in ("auto", "bf16", "bfloat16"):
+        raise ValueError(
+            "Learnable DSA attention sinks require a bfloat16 KV cache; "
+            f"got kv_cache_dtype={kv_cache_dtype}."
+        )
     if kv_cache_dtype == "auto":
-        kv_cache_dtype = "fp8_e4m3" if major >= 10 else "bfloat16"
+        kv_cache_dtype = (
+            "fp8_e4m3" if major >= 10 and not has_attention_sinks else "bfloat16"
+        )
         logger.warning(
             f"Setting KV cache dtype to {kv_cache_dtype} for DeepSeek DSA on SM{major} device."
         )
@@ -697,6 +705,27 @@ def _dsa_split_backend_resolution(view: Any) -> dict:
         and not get_platform().is_hip
     )
 
+    if getattr(hf_config, "learnable_sink", False):
+        backend = "flashmla_sparse"
+        for field in ("dsa_prefill_backend", "dsa_decode_backend"):
+            value = getattr(view, field)
+            if value is not None and value != backend:
+                option = "--" + field.replace("_", "-")
+                raise ValueError(
+                    f"{model_arch} uses learnable attention sinks and requires "
+                    f"{option} {backend!r}; got {value!r}"
+                )
+        if not user_set_prefill:
+            declared["dsa_prefill_backend"] = backend
+        if not user_set_decode:
+            declared["dsa_decode_backend"] = backend
+        logger.warning(
+            "Set DSA backends for learnable attention sinks: "
+            f"prefill={declared.get('dsa_prefill_backend', view.dsa_prefill_backend)}, "
+            f"decode={declared.get('dsa_decode_backend', view.dsa_decode_backend)}."
+        )
+        return declared
+
     if is_glm_sm12_fp8:
         backend = "flashinfer_sparse_mla"
         if not user_set_prefill:
@@ -763,6 +792,8 @@ _DEEPSEEK_FAMILY_ARCHS = frozenset(
         "MistralLarge3ForCausalLM",
         "PixtralForConditionalGeneration",
         "GlmMoeDsaForCausalLM",
+        "HYV4ForCausalLM",
+        "HYV4ForCausalLMNextN",
         "LongcatFlashForCausalLM",
         "LongcatFlashForCausalLMNextN",
         "Dots3NoteForCausalLM",

python/sglang/srt/arg_groups/speculative_hook.py

@@ -862,6 +862,7 @@ def _handle_eagle_family(server_args: ServerArgs) -> None:
         "MistralLarge3ForCausalLM",
         "PixtralForConditionalGeneration",
         "HYV3ForCausalLM",
+        "HYV4ForCausalLM",
     ]:
         if cfg.speculative_draft_model_path is None:
             declare_resolution(

python/sglang/srt/configs/__init__.py

@@ -17,6 +17,7 @@ from sglang.srt.configs.dots_vlm import DotsVLMConfig
 from sglang.srt.configs.exaone import ExaoneConfig
 from sglang.srt.configs.falcon_h1 import FalconH1Config
 from sglang.srt.configs.granitemoehybrid import GraniteMoeHybridConfig
+from sglang.srt.configs.hy_v4 import HYV4Config
 from sglang.srt.configs.inkling import (
     InklingAudioConfig,
     InklingMMConfig,
@@ -118,6 +119,7 @@ __all__ = [
     "Dots3Config",
     "FalconH1Config",
     "GraniteMoeHybridConfig",
+    "HYV4Config",
     "Lfm2Config",
     "Lfm2MoeConfig",
     "Lfm2VlConfig",

python/sglang/srt/configs/hy_v4.py

@@ -0,0 +1,152 @@
+from transformers.configuration_utils import PreTrainedConfig
+
+
+class HYV4Config(PreTrainedConfig):
+    model_type = "hy_v4"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_local_experts": "n_routed_experts"}
+
+    def __init__(
+        self,
+        vocab_size=120832,
+        hidden_size=2816,
+        intermediate_size=6912,
+        moe_intermediate_size=768,
+        num_hidden_layers=34,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=262144,
+        rms_norm_eps=1e-5,
+        n_routed_experts=256,
+        n_shared_experts=1,
+        num_experts_per_tok=8,
+        routed_scaling_factor=2.827,
+        norm_topk_prob=True,
+        q_lora_rank=1536,
+        kv_lora_rank=512,
+        qk_nope_head_dim=192,
+        qk_rope_head_dim=64,
+        v_head_dim=256,
+        mlp_layer_types=None,
+        layer_types=None,
+        index_topk=2048,
+        index_head_dim=128,
+        index_n_heads=16,
+        indexer_types=None,
+        enable_lm_head_fp32=True,
+        enable_ihc=True,
+        hc_mult=4,
+        hc_magnitude=2.0,
+        hc_eps=1e-6,
+        gated_mla=True,
+        gating_type="elementwise",
+        learnable_sink=True,
+        learnable_sink_init=0.0,
+        swiglu_limit=10.0,
+        rope_parameters=None,
+        num_nextn_predict_layers=1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.rms_norm_eps = rms_norm_eps
+        self.n_routed_experts = n_routed_experts
+        self.n_shared_experts = n_shared_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.routed_scaling_factor = routed_scaling_factor
+        self.norm_topk_prob = norm_topk_prob
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.index_topk = index_topk
+        self.index_head_dim = index_head_dim
+        self.index_n_heads = index_n_heads
+        self.enable_lm_head_fp32 = enable_lm_head_fp32
+        self.enable_ihc = enable_ihc
+        self.hc_mult = hc_mult
+        self.hc_magnitude = hc_magnitude
+        self.hc_eps = hc_eps
+        self.gated_mla = gated_mla
+        self.gating_type = gating_type
+        self.learnable_sink = learnable_sink
+        self.learnable_sink_init = learnable_sink_init
+        self.swiglu_limit = swiglu_limit
+        self.rope_parameters = (
+            rope_parameters
+            if rope_parameters is not None
+            else {"rope_theta": 10000000.0, "rope_type": "default"}
+        )
+        self.rope_theta = self.rope_parameters["rope_theta"]
+        self.rope_interleave = True
+        self.indexer_rope_interleave = True
+        self.router_fp32 = True
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.head_dim = qk_rope_head_dim
+        self.mlp_layer_types = (
+            mlp_layer_types
+            if mlp_layer_types is not None
+            else ["dense"] + ["sparse"] * (num_hidden_layers - 1)
+        )
+        self.layer_types = (
+            layer_types
+            if layer_types is not None
+            else ["deepseek_sparse_attention"] * num_hidden_layers
+        )
+        self.indexer_types = (
+            indexer_types
+            if indexer_types is not None
+            else [
+                "full" if i == 0 or (i - 1) % 4 == 0 else "shared"
+                for i in range(num_hidden_layers)
+            ]
+        )
+        self.first_k_dense_replace = 1
+        self.moe_layer_freq = 1
+        self.scoring_func = "sigmoid"
+        self.topk_method = "noaux_tc"
+        self.n_group = 1
+        self.topk_group = 1
+        self._validate_hy_v4()
+
+    def _validate_hy_v4(self):
+        fields = {
+            "mlp_layer_types": self.mlp_layer_types,
+            "layer_types": self.layer_types,
+            "indexer_types": self.indexer_types,
+        }
+        for name, values in fields.items():
+            if len(values) != self.num_hidden_layers:
+                raise ValueError(
+                    f"{name} must contain {self.num_hidden_layers} entries, got {len(values)}"
+                )
+        if set(self.mlp_layer_types) - {"dense", "sparse"}:
+            raise ValueError("mlp_layer_types only supports dense and sparse")
+        if set(self.layer_types) != {"deepseek_sparse_attention"}:
+            raise ValueError("HYV4 only supports deepseek_sparse_attention")
+        if set(self.indexer_types) - {"full", "shared"}:
+            raise ValueError("indexer_types only supports full and shared")
+        if self.indexer_types[0] != "full":
+            raise ValueError("indexer_types must start with a full indexer")
+        if not self.enable_ihc or self.hc_mult <= 0:
+            raise ValueError("HYV4 requires enabled iHC with hc_mult > 0")
+        if not self.gated_mla or self.gating_type != "elementwise":
+            raise ValueError("HYV4 requires elementwise gated MLA")
+        if not self.learnable_sink:
+            raise ValueError("HYV4 requires learnable attention sinks")
+        if self.q_lora_rank is None:
+            raise ValueError("HYV4 sparse attention requires q_lora_rank")
+
+
+__all__ = ["HYV4Config"]

python/sglang/srt/configs/model_config.py

@@ -66,6 +66,14 @@ def _quant_config_to_dict(quant_config):
     return quant_config
 
 
+def unwrap_modelopt_quantization_config(quant_config: dict) -> dict:
+    quantization = quant_config.get("quantization", quant_config)
+    if not isinstance(quantization, dict):
+        return {}
+    nested = quantization.get("quantization")
+    return nested if isinstance(nested, dict) else quantization
+
+
 def get_mimo_v2_fused_qkv_expected_tp_size(hf_config):
     layout = getattr(hf_config, "attention_projection_layout", None)
     if layout is None:
@@ -134,6 +142,8 @@ def is_deepseek_dsa(config) -> bool:
             "LongcatFlashForCausalLMNextN",
             "Dots3NoteForCausalLM",
             "Dots3NoteForCausalLMNextN",
+            "HYV4ForCausalLM",
+            "HYV4ForCausalLMNextN",
         )
         and _hf_attr(config, "index_topk") is not None
     )
@@ -167,6 +177,17 @@ def is_deepseek_v4(config) -> bool:
     )
 
 
+def resolve_spec_hidden_size(
+    hf_config, hidden_size: int, hc_mult: int
+) -> tuple[int, Optional[int]]:
+    # Only DSV4 carries the hc-flattened stream across the target→draft
+    # boundary; other hc models (hy_v4) collapse to hidden_size first.
+    if hc_mult <= 1 or not is_deepseek_v4(hf_config):
+        return hidden_size, None
+    hc_hidden_size = hidden_size * hc_mult
+    return hc_hidden_size, hc_hidden_size
+
+
 def get_dsa_index_head_dim(config: PretrainedConfig) -> int:
     assert is_deepseek_dsa(config) or is_deepseek_v4(config)
     return config.index_head_dim
@@ -226,6 +247,12 @@ def dsa_layer_skips_topk(config: PretrainedConfig, layer_id: int) -> bool:
     """Return whether a DSA layer reuses the previous layer's top-k indices."""
     assert is_deepseek_dsa(config)
 
+    indexer_types = getattr(config, "indexer_types", None)
+    if indexer_types is not None:
+        return (
+            0 <= layer_id < len(indexer_types) and indexer_types[layer_id] == "shared"
+        )
+
     # LongCat computes fresh top-k indices every cli_factor layers.
     cli_factor = getattr(config, "cli_factor", 1)
     if cli_factor is None:
@@ -794,6 +821,10 @@ class ModelConfig:
             self.hf_config.architectures[0] = "HYV3ForCausalLMNextN"
             self.hf_config.num_nextn_predict_layers = 1
 
+        if is_draft_model and self.hf_config.architectures[0] == "HYV4ForCausalLM":
+            self.hf_config.architectures[0] = "HYV4ForCausalLMNextN"
+            self.hf_config.num_nextn_predict_layers = 1
+
     def _derive_hybrid_model(self):
         # Use self.context_len after it has been initialized to prevent using context_len which may be None.
         self.is_hybrid_swa = (
@@ -865,7 +896,7 @@ class ModelConfig:
             return getattr(
                 self.hf_text_config, "add_swa_attention_sink_bias", False
             ) or getattr(self.hf_text_config, "add_full_attention_sink_bias", False)
-        return False
+        return bool(getattr(self.hf_text_config, "learnable_sink", False))
 
     def _derive_context_length(self, context_length: int):
         is_draft_model = self.is_draft_model
@@ -939,6 +970,8 @@ class ModelConfig:
             or "GlmMoeDsaForCausalLMNextN" in self.hf_config.architectures
             or "LongcatFlashForCausalLM" in self.hf_config.architectures
             or "LongcatFlashForCausalLMNextN" in self.hf_config.architectures
+            or "HYV4ForCausalLM" in self.hf_config.architectures
+            or "HYV4ForCausalLMNextN" in self.hf_config.architectures
             or "DotsVLMForCausalLM" in self.hf_config.architectures
             or "Dots3NoteForCausalLM" in self.hf_config.architectures
             or "Dots3NoteForCausalLMNextN" in self.hf_config.architectures
@@ -969,7 +1002,10 @@ class ModelConfig:
                 else None
             )
             # In transformers v5, rope_scaling is just rope_parameters.
-            self._init_mla_scaling(self.hf_text_config.rope_scaling)
+            rope_scaling = getattr(
+                self.hf_text_config, "rope_parameters", None
+            ) or getattr(self.hf_text_config, "rope_scaling", None)
+            self._init_mla_scaling(rope_scaling)
         elif (
             "DeepseekV4ForCausalLM" in self.hf_config.architectures
             or "DeepseekV4ForCausalLMNextN" in self.hf_config.architectures
@@ -1102,12 +1138,9 @@ class ModelConfig:
             self.num_key_value_heads = self.num_attention_heads
         self.hidden_size = self.hf_text_config.hidden_size
         hc_mult = getattr(self.hf_text_config, "hc_mult", 1)
-        self.spec_hidden_size = (
-            self.hidden_size * hc_mult if hc_mult > 1 else self.hidden_size
+        self.spec_hidden_size, self.hc_hidden_size = resolve_spec_hidden_size(
+            self.hf_config, self.hidden_size, hc_mult
         )
-        # mHC-flattened hidden size; None when not running an mHC model
-        # (e.g. non-DeepSeek-V4 configs without ``hc_mult``).
-        self.hc_hidden_size = self.spec_hidden_size if hc_mult > 1 else None
         self.num_hidden_layers = self.hf_text_config.num_hidden_layers
         self.num_attention_layers = self.num_hidden_layers
         if "LongcatFlashForCausalLM" in self.hf_config.architectures:
@@ -1370,11 +1403,8 @@ class ModelConfig:
 
             {"quant_algo": "FP8", "quant_method": "modelopt", ...}
         """
-        if "quantization" in quant_config_dict:
-            json_quant_configs = quant_config_dict["quantization"]
-        elif "quant_algo" in quant_config_dict:
-            json_quant_configs = quant_config_dict
-        else:
+        json_quant_configs = unwrap_modelopt_quantization_config(quant_config_dict)
+        if not json_quant_configs:
             return None
         quant_algo = json_quant_configs.get("quant_algo", None)
 

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -89,6 +89,10 @@ from sglang.srt.function_call.utils import (
 )
 from sglang.srt.managers.io_struct import GenerateReqInput
 from sglang.srt.parser.conversation import generate_chat_conv
+from sglang.srt.parser.hunyuan_reasoning import (
+    normalize_hunyuan_reasoning_effort,
+    uses_hunyuan_reasoning_effort,
+)
 from sglang.srt.parser.jinja_template_utils import (
     MEDIA_URL_PART_TYPES,
     process_content_for_template_format,
@@ -1048,6 +1052,7 @@ class OpenAIServingChat(OpenAIServingBase):
         request: ChatCompletionRequest,
         raw_request: Request = None,
     ) -> tuple[GenerateReqInput, ChatCompletionRequest]:
+
         # Header-based opt-in (same rationale as request_headers.py).
         if raw_request is not None and not request.return_input_ids_in_sglext:
             if raw_request.headers.get("x-sglext-return-input-ids") == "1":
@@ -1057,11 +1062,16 @@ class OpenAIServingChat(OpenAIServingBase):
             if raw_request.headers.get("x-sglext-return-output-ids") == "1":
                 request.return_output_ids_in_sglext = True
 
-        reasoning_effort = (
-            request.chat_template_kwargs.pop("reasoning_effort", None)
-            if request.chat_template_kwargs
-            else None
-        )
+        reasoning_effort = None
+        if not uses_hunyuan_reasoning_effort(
+            self.reasoning_parser, self.template_manager.reasoning_config
+        ):
+            reasoning_effort = (
+                request.chat_template_kwargs.pop("reasoning_effort", None)
+                if request.chat_template_kwargs
+                else None
+            )
+
         if self.is_gpt_oss and reasoning_effort == "none":
             raise ValueError(
                 f"Harmony does not support reasoning effort {reasoning_effort}"
@@ -1205,6 +1215,10 @@ class OpenAIServingChat(OpenAIServingBase):
             if effort is not None and request.reasoning_effort is None:
                 request.reasoning_effort = effort
 
+        normalize_hunyuan_reasoning_effort(
+            request, self.reasoning_parser, self.template_manager.reasoning_config
+        )
+
         # GptOss model needs to keep special tokens for harmony parsing
         if self.is_gpt_oss or self.is_gemma4:
             request.skip_special_tokens = False
@@ -2575,7 +2589,11 @@ class OpenAIServingChat(OpenAIServingBase):
             return
 
         if self.reasoning_parser == "hunyuan":
-            request.reasoning_effort = "medium" if enabled else "no_think"
+            config = self.template_manager.reasoning_config
+            if config is not None and config.special_case == "hunyuan_effort":
+                request.reasoning_effort = "high" if enabled else "no_think"
+            else:
+                request.reasoning_effort = "medium" if enabled else "no_think"
             return
 
         if self.reasoning_parser == "inkling":
@@ -2644,6 +2662,9 @@ class OpenAIServingChat(OpenAIServingBase):
             ) == "enabled"
 
         if self.reasoning_parser == "hunyuan":
+            config = self.template_manager.reasoning_config
+            if config is not None and config.special_case == "hunyuan_effort":
+                return request.reasoning_effort not in ("none", "no_think")
             # Hy3-preview template emits no <think> when reasoning_effort is
             # "no_think" / "none" / unset; forcing reasoning would route all
             # output into reasoning_content.

python/sglang/srt/function_call/hunyuan_detector.py

@@ -103,6 +103,20 @@ class HunyuanDetector(BaseFormatDetector):
     def __init__(self, tokenizer=None):
         super().__init__()
 
+        # Hy4 dropped <tool_sep> from its vocab entirely; Hy3 carries it (bare
+        # or suffixed). Detect the dialect from the vocab, not from the
+        # resolved literal (which falls back to "<tool_sep>" for both).
+        self._has_tool_sep = True
+        if tokenizer is not None:
+            try:
+                self._has_tool_sep = any(
+                    re.fullmatch(r"<tool_sep(?::[^>]+)?>", tok)
+                    for tok in tokenizer.get_vocab()
+                    if isinstance(tok, str)
+                )
+            except Exception:
+                pass
+
         t = resolve_hunyuan_tokens(tokenizer)
         tool_calls = t["tool_calls"]
         tool_call = t["tool_call"]
@@ -129,7 +143,13 @@ class HunyuanDetector(BaseFormatDetector):
         self.tool_call_regex = re.compile(
             re.escape(tool_call)
             + r"(.*?)"
+            + r"(?:"
             + re.escape(tool_sep)
+            + r"|(?="
+            + re.escape(arg_key)
+            + r"|"
+            + re.escape(tc_end)
+            + r"))"
             + r"(.*?)"
             + re.escape(tc_end),
             re.DOTALL,
@@ -383,14 +403,22 @@ class HunyuanDetector(BaseFormatDetector):
                         self._in_tool_calls = False
                     break
 
-                sep_pos = self._buffer.find(self.tool_sep_token, tc_start)
-                if sep_pos == -1:
+                name_start = tc_start + len(self.tool_call_start_token)
+                boundaries = [
+                    (self._buffer.find(token, name_start), token)
+                    for token in (
+                        self.tool_sep_token,
+                        self.arg_key_start_token,
+                        self.tool_call_end_token,
+                    )
+                ]
+                boundaries = [(pos, token) for pos, token in boundaries if pos != -1]
+                if not boundaries:
                     self._buffer = self._buffer[tc_start:]
                     break
+                boundary_pos, boundary_token = min(boundaries, key=lambda item: item[0])
 
-                tool_name = self._buffer[
-                    tc_start + len(self.tool_call_start_token) : sep_pos
-                ].strip()
+                tool_name = self._buffer[name_start:boundary_pos].strip()
 
                 if (
                     tool_name not in self._tool_indices
@@ -413,7 +441,9 @@ class HunyuanDetector(BaseFormatDetector):
                     )
                 )
 
-                self._buffer = self._buffer[sep_pos + len(self.tool_sep_token) :]
+                if boundary_token == self.tool_sep_token:
+                    boundary_pos += len(boundary_token)
+                self._buffer = self._buffer[boundary_pos:]
 
             # Phase 2: stream argument JSON of the current tool.
             before_name = self._streaming_tool_name
@@ -534,6 +564,14 @@ class HunyuanDetector(BaseFormatDetector):
         return []
 
     def structure_info(self) -> _GetInfoFunc:
+        # Hy4 has no <tool_sep> and no separators between adjacent
+        # <tool_call> blocks, so the Hy3 begin/end literals never match.
+        if not self._has_tool_sep:
+            return lambda name: StructureInfo(
+                begin=f"{self.bot_token}{self.tool_call_start_token}{name}",
+                end=f"{self.tool_call_end_token}{self.eot_token}",
+                trigger=self.bot_token,
+            )
         return lambda name: StructureInfo(
             begin=f"{self.bot_token}\n{self.tool_call_start_token}{name}{self.tool_sep_token}",
             end=f"{self.tool_call_end_token}\n{self.eot_token}",
@@ -542,3 +580,9 @@ class HunyuanDetector(BaseFormatDetector):
 
     def supports_structural_tag(self) -> bool:
         return False
+
+    def parses_required_natively(self) -> bool:
+        # Hy3/Hy4 both emit their structural tool-call format natively; without
+        # a dialect grammar, a json_schema constraint would force plain JSON the
+        # model was not trained to produce here.
+        return True

python/sglang/srt/layers/attention/dsa_backend.py

@@ -348,6 +348,7 @@ class DeepseekSparseAttnBackend(
             # Keep original head count if it exceeds current padded variants.
             self.flashmla_kv_num_q_heads = self.num_q_heads
         self.enable_auto_select_prefill_impl = self.dsa_prefill_impl == "flashmla_auto"
+        self._sink_pad_cache: dict[tuple[int, int], torch.Tensor] = {}
 
         self._arange_buf = torch.arange(16384, device=self.device, dtype=torch.int32)
 
@@ -1885,6 +1886,7 @@ class DeepseekSparseAttnBackend(
         cos_sin_cache: Optional[torch.Tensor] = None,
         is_neox: Optional[bool] = False,
         llama_4_scaling: Optional[torch.Tensor] = None,
+        attn_sink: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
 
         causal = not layer.is_cross_attention
@@ -1899,6 +1901,10 @@ class DeepseekSparseAttnBackend(
             )
             else self.dsa_prefill_impl
         )
+        if attn_sink is not None and dsa_impl != "flashmla_sparse":
+            raise RuntimeError(
+                f"Learnable attention sinks require flashmla_sparse, got {dsa_impl}"
+            )
 
         if dsa_impl == "trtllm" and not self.use_mha:
             return self._forward_trtllm(
@@ -2124,6 +2130,7 @@ class DeepseekSparseAttnBackend(
                 sm_scale=layer.scaling,
                 v_head_dim=layer.v_head_dim,
                 topk_length=metadata.dsa_cache_seqlens_int32,
+                attn_sink=attn_sink,
             )
         elif dsa_impl == "flashinfer_sparse_mla":
             if q_rope is not None:
@@ -2195,12 +2202,19 @@ class DeepseekSparseAttnBackend(
         cos_sin_cache: Optional[torch.Tensor] = None,
         is_neox: Optional[bool] = False,
         llama_4_scaling: Optional[torch.Tensor] = None,
+        attn_sink: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
 
         causal = not layer.is_cross_attention
         metadata = self.forward_metadata
         assert causal, "DSA is causal only"
 
+        if attn_sink is not None and self.dsa_decode_impl != "flashmla_sparse":
+            raise RuntimeError(
+                "Learnable attention sinks require flashmla_sparse, got "
+                f"{self.dsa_decode_impl}"
+            )
+
         if self.dsa_decode_impl == "trtllm":
             return self._forward_trtllm(
                 q,
@@ -2281,6 +2295,7 @@ class DeepseekSparseAttnBackend(
                 sm_scale=layer.scaling,
                 v_head_dim=layer.v_head_dim,
                 topk_length=metadata.dsa_cache_seqlens_int32,
+                attn_sink=attn_sink,
             )
         elif self.dsa_decode_impl == "flashinfer_sparse_mla":
             if q_all is None:
@@ -2396,6 +2411,7 @@ class DeepseekSparseAttnBackend(
         page_table_1: torch.Tensor,
         sm_scale: float,
         topk_length: Optional[torch.Tensor] = None,
+        attn_sink: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         from sgl_kernel.flash_mla import flash_mla_sparse_fwd
 
@@ -2421,6 +2437,15 @@ class DeepseekSparseAttnBackend(
         else:
             q_input = q_all
 
+        sink_input = attn_sink
+        if need_padding and attn_sink is not None:
+            key = (attn_sink.data_ptr(), required_padding)
+            sink_input = self._sink_pad_cache.get(key)
+            if sink_input is None:
+                sink_input = attn_sink.new_zeros(required_padding)
+                self._sink_pad_cache[key] = sink_input
+            sink_input[:num_heads].copy_(attn_sink)
+
         # indices shape must be (s_q, h_kv=1, topk), keep h_kv=1 unchanged
         indices_input = page_table_1.unsqueeze(1)
 
@@ -2442,6 +2467,7 @@ class DeepseekSparseAttnBackend(
             indices=indices_input,
             sm_scale=sm_scale,
             d_v=v_head_dim,
+            attn_sink=sink_input,
             topk_length=topk_length,
         )
 
@@ -3437,6 +3463,7 @@ class DeepseekSparseAttnMultiStepBackend:
     ):
         self.topk = topk
         self.speculative_num_steps = speculative_num_steps
+        self._sink_pad_cache: dict[tuple[int, int], torch.Tensor] = {}
         self.attn_backends = []
         for i in range(self.speculative_num_steps - 1):
             self.attn_backends.append(

python/sglang/srt/layers/deep_gemm_wrapper/entrypoint.py

@@ -170,6 +170,52 @@ def grouped_gemm_nt_bf16_contig(
         deep_gemm.m_grouped_bf16_gemm_nt_contiguous(a, b, d, m_indices)
 
 
+def get_contiguous_layout_alignment(expected_m: int, num_groups: int) -> int:
+    alignment = deep_gemm.get_mk_alignment_for_contiguous_layout()
+    if isinstance(alignment, tuple):
+        alignment = alignment[0]
+
+    theoretical_alignment = getattr(
+        deep_gemm, "get_theoretical_mk_alignment_for_contiguous_layout", None
+    )
+    if theoretical_alignment is None:
+        return alignment
+
+    per_group_m = (expected_m + num_groups - 1) // num_groups
+    try:
+        try:
+            candidate = theoretical_alignment(per_group_m)
+        except TypeError:
+            candidate = theoretical_alignment(
+                expected_m=expected_m, num_groups=num_groups
+            )
+    except Exception:
+        return alignment
+    return candidate if 0 < candidate <= alignment else alignment
+
+
+@contextmanager
+def contiguous_layout_alignment_scope(alignment: Optional[int]):
+    if alignment is None:
+        yield
+        return
+
+    getter = getattr(deep_gemm, "get_mk_alignment_for_contiguous_layout", None)
+    setter = getattr(deep_gemm, "set_mk_alignment_for_contiguous_layout", None)
+    if getter is None or setter is None:
+        yield
+        return
+
+    previous = getter()
+    if isinstance(previous, tuple):
+        previous = previous[0]
+    setter(alignment)
+    try:
+        yield
+    finally:
+        setter(previous)
+
+
 def gemm_nt_f8f8bf16(
     lhs: Tuple[torch.Tensor, torch.Tensor],
     rhs: Tuple[torch.Tensor, torch.Tensor],

python/sglang/srt/layers/logits_processor.py

@@ -386,7 +386,9 @@ class LogitsProcessor(nn.Module):
         self.logit_scale = logit_scale
         self.use_attn_tp_group = get_parallel().enable_dp_lm_head
         self.use_tp_lm_head_all_to_all = get_parallel().enable_tp_lm_head_all_to_all
-        self.use_fp32_lm_head = get_exec().features.enable_fp32_lm_head
+        self.use_fp32_lm_head = get_exec().features.enable_fp32_lm_head or getattr(
+            config, "enable_lm_head_fp32", False
+        )
         if self.use_attn_tp_group:
             self.attn_tp_size = get_parallel().attn_tp_size
             self.do_tensor_parallel_all_gather = (

python/sglang/srt/layers/moe/cutlass_moe.py

@@ -50,6 +50,7 @@ def cutlass_fused_experts_fp8(
     use_mxfp8: bool = False,
     output: Optional[torch.Tensor] = None,
     enable_es: Tuple[bool, bool] = (False, False),
+    swiglu_limit: Optional[float] = None,
 ) -> torch.Tensor:
     """Performs Fused MoE computation using CUTLASS-like kernels with FP8 weights and activations.
 
@@ -271,7 +272,12 @@ def cutlass_fused_experts_fp8(
         )
 
     intermediate = torch.empty((m * topk, n), device=device, dtype=out_dtype)
-    silu_and_mul(c1, intermediate)
+    if swiglu_limit is None:
+        silu_and_mul(c1, intermediate)
+    else:
+        from sglang.kernels.ops.attention.dsv4 import silu_and_mul_clamp
+
+        silu_and_mul_clamp(c1, intermediate, swiglu_limit)
 
     if use_mxfp8 and es_down:
         intemediate_q = torch.empty_like(intermediate, dtype=torch.float8_e4m3fn)

python/sglang/srt/layers/moe/moe_runner/deep_gemm.py

@@ -8,7 +8,11 @@ import torch
 import triton
 import triton.language as tl
 
-from sglang.kernels.ops.attention.dsv4 import silu_and_mul_masked_post_quant
+from sglang.kernels.ops.attention.dsv4 import (
+    silu_and_mul_clamp,
+    silu_and_mul_masked_post_quant,
+)
+from sglang.kernels.ops.moe.triton_pad_expert_counts import pad_expert_counts
 from sglang.kernels.ops.quantization import per_token_group_quant
 
 logger = logging.getLogger(__name__)
@@ -164,6 +168,32 @@ def _estimate_masked_standard_layout_peak_bytes(
     return runner_config.num_local_experts * padded_m * peak_row_bytes
 
 
+_masked_activation_fallback_logged = False
+
+
+# Masked clamped/swizzled activation only exists as the DSV4 JIT kernel, which
+# requires D // 8 >= E and group 128 (silu_and_mul_masked_post_quant.cuh:245).
+def _masked_activation_unsupported_reason(
+    runner_config: MoeRunnerConfig, quant_info: DeepGemmMoeQuantInfo
+) -> Optional[str]:
+    if runner_config.swiglu_limit is None and not get_moe_a2a_backend().is_megamoe():
+        return None
+    d = runner_config.intermediate_size_per_partition
+    e = runner_config.num_local_experts
+    if d is None or e is None:
+        return None
+    group_size = quant_info.block_shape[1] if quant_info.block_shape else 128
+    if d // 8 < e:
+        return f"D // 8 ({d // 8}) < num_local_experts ({e})"
+    if group_size != 128:
+        return (
+            f"masked activation group_size {group_size}, DSV4 JIT kernel requires 128"
+        )
+    if d % (group_size * 4) != 0:
+        return f"D ({d}) not divisible by 4 * group_size"
+    return None
+
+
 def _should_use_masked_standard_layout(
     runner_config: MoeRunnerConfig,
     quant_info: DeepGemmMoeQuantInfo,
@@ -177,6 +207,18 @@ def _should_use_masked_standard_layout(
     ):
         return False
 
+    reason = _masked_activation_unsupported_reason(runner_config, quant_info)
+    if reason is not None:
+        global _masked_activation_fallback_logged
+        if not _masked_activation_fallback_logged:
+            _masked_activation_fallback_logged = True
+            logger.info(
+                "DeepGEMM masked standard layout disabled: %s. "
+                "Clamped/swizzled activations on this config must use the "
+                "compact layout.",
+                reason,
+            )
+        return False
     mode = envs.SGLANG_DEEPGEMM_STANDARD_LAYOUT.get().lower()
     if mode not in ("auto", "masked", "compact"):
         raise ValueError(
@@ -287,24 +329,30 @@ class DeepGemmRunnerCore(MoeRunnerCore):
         hooks: Optional[Any] = None,
     ) -> DeepGemmRunnerOutput:
         weight_dtype = quant_info.w13_weight.dtype
-        if not runner_input.use_masked_gemm:
-            if weight_dtype == torch.bfloat16:
-                hidden_states = self._run_bf16_contiguous_gemm(
-                    runner_input, quant_info, running_state
-                )
+        alignment = (
+            running_state.get("contiguous_layout_alignment")
+            if not runner_input.use_masked_gemm
+            else None
+        )
+        with deep_gemm_wrapper.contiguous_layout_alignment_scope(alignment):
+            if not runner_input.use_masked_gemm:
+                if weight_dtype == torch.bfloat16:
+                    hidden_states = self._run_bf16_contiguous_gemm(
+                        runner_input, quant_info, running_state
+                    )
+                else:
+                    hidden_states = self._run_contiguous_gemm(
+                        runner_input, quant_info, running_state
+                    )
             else:
-                hidden_states = self._run_contiguous_gemm(
-                    runner_input, quant_info, running_state
-                )
-        else:
-            if weight_dtype == torch.bfloat16:
-                hidden_states = self._run_masked_bf16_gemm(
-                    runner_input, quant_info, running_state
-                )
-            else:
-                hidden_states = self._run_masked_gemm(
-                    runner_input, quant_info, running_state
-                )
+                if weight_dtype == torch.bfloat16:
+                    hidden_states = self._run_masked_bf16_gemm(
+                        runner_input, quant_info, running_state
+                    )
+                else:
+                    hidden_states = self._run_masked_gemm(
+                        runner_input, quant_info, running_state
+                    )
         return DeepGemmRunnerOutput(hidden_states=hidden_states)
 
     def _run_contiguous_gemm(
@@ -333,7 +381,7 @@ class DeepGemmRunnerCore(MoeRunnerCore):
 
         N = quant_info.w13_weight.size(1)
         K = hidden_states_shape[1]
-        scale_block_size = 128
+        scale_block_size = quant_info.block_shape[1] if quant_info.use_mxfp8 else 128
 
         if all_tokens == 0:
             if trace_deepep_v2_contig:
@@ -344,9 +392,12 @@ class DeepGemmRunnerCore(MoeRunnerCore):
                 (0, K), device=hidden_states_device, dtype=torch.bfloat16
             )
 
-        recipe_a, recipe_b = (
-            ((1, 128), (1, 32)) if quant_info.is_fp4_experts else (None, None)
-        )
+        if quant_info.use_mxfp8:
+            recipe_a = recipe_b = tuple(quant_info.block_shape)
+        elif quant_info.is_fp4_experts:
+            recipe_a, recipe_b = (1, 128), (1, 32)
+        else:
+            recipe_a, recipe_b = None, None
 
         w13_weight_fp8 = (
             quant_info.w13_weight,
@@ -469,19 +520,25 @@ class DeepGemmRunnerCore(MoeRunnerCore):
                 sglang_per_token_group_quant_fp8,
             )
 
-            if self.swiglu_limit is not None:
-                gateup_output = _apply_swiglu_limit(
-                    gateup_output, swiglu_limit=self.swiglu_limit
-                )
-
             if not _is_musa:
                 down_input = torch.empty(
                     (all_tokens, N // 2),
                     device=gateup_output.device,
                     dtype=torch.bfloat16,
                 )
-                _legacy_silu_and_mul(gateup_output.view(-1, N), down_input)
+                if self.swiglu_limit is not None:
+                    # Fuse the SwiGLU limit with the activation. The quantizing
+                    # sibling only supports a group size of 128.
+                    silu_and_mul_clamp(
+                        gateup_output.view(-1, N), down_input, self.swiglu_limit
+                    )
+                else:
+                    _legacy_silu_and_mul(gateup_output.view(-1, N), down_input)
             else:
+                if self.swiglu_limit is not None:
+                    gateup_output = _apply_swiglu_limit(
+                        gateup_output, swiglu_limit=self.swiglu_limit
+                    )
                 down_input = _silu_and_mul_musa(gateup_output.view(-1, N))
             del gateup_output
 
@@ -512,12 +569,17 @@ class DeepGemmRunnerCore(MoeRunnerCore):
         if deep_gemm_wrapper.DEEPGEMM_NEED_TMA_ALIGNED_SCALES:
             down_input_scale = tma_align_input_scale(down_input_scale)
 
+        recipe_a_down = (
+            (quant_info.block_shape[0], scale_block_size)
+            if quant_info.use_mxfp8
+            else recipe_a
+        )
         deep_gemm_wrapper.grouped_gemm_nt_f8f8bf16_contig(
             (down_input_fp8, down_input_scale),
             w2_weight_fp8,
             down_output,
             m_indices,
-            recipe_a=recipe_a,
+            recipe_a=recipe_a_down,
             recipe_b=recipe_b,
         )
         if trace_deepep_v2_contig:
@@ -976,9 +1038,11 @@ def pre_permute_standard_to_deep_gemm(
 
     # The compact layout avoids scaling masked buffers with the expert count.
     # Scatter and post-permute skip non-local experts mapped to -1.
-    block_e = 128
     num_experts = runner_config.num_local_experts
     num_assignments = topk_ids.numel()
+    block_e = deep_gemm_wrapper.get_contiguous_layout_alignment(
+        num_assignments, num_experts
+    )
     all_tokens = _get_compact_all_tokens(num_assignments, num_experts, block_e)
 
     tokens_per_expert, unused_masked_dst = fused_moe_dispatch_index(
@@ -986,10 +1050,15 @@ def pre_permute_standard_to_deep_gemm(
     )
     dispose_tensor(unused_masked_dst)
     valid_tokens_per_expert = tokens_per_expert
-    tokens_per_expert = (ceil_div(tokens_per_expert, block_e) * block_e).to(torch.int32)
-    # Keep graph-static shapes by appending padding to the final segment.
-    # Its m_indices stay -1, so DeepGEMM skips those rows.
-    tokens_per_expert[-1].add_(all_tokens - tokens_per_expert.sum())
+    if _is_cuda:
+        tokens_per_expert = pad_expert_counts(tokens_per_expert, block_e, all_tokens)
+    else:
+        # The Triton kernel is CUDA-only. Keep the existing MUSA-compatible
+        # tensor implementation for other DeepGEMM backends.
+        tokens_per_expert = (ceil_div(tokens_per_expert, block_e) * block_e).to(
+            torch.int32
+        )
+        tokens_per_expert[-1].add_(all_tokens - tokens_per_expert.sum())
 
     k = hidden_states.size(1)
     output_dtype = (
@@ -1023,7 +1092,13 @@ def pre_permute_standard_to_deep_gemm(
                 scale_ue8m0=deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0,
             )
         )
-        packed_input = torch.zeros(
+        # ep_scatter writes every live row and the grouped GEMM's results for
+        # the alignment padding are dropped by post_reorder, so the zeroing is
+        # dead work -- 174 MB per layer at bs=64. The sibling dispatch path in
+        # this file already allocates its equivalent buffer with torch.empty
+        # unless deterministic inference is on; match it.
+        deterministic = get_exec().deterministic.enable_deterministic_inference
+        packed_input = (torch.zeros if deterministic else torch.empty)(
             (all_tokens, k),
             device=hidden_states_device,
             dtype=torch.float8_e4m3fn,
@@ -1062,6 +1137,7 @@ def pre_permute_standard_to_deep_gemm(
         src2dst,
         scale_ue8m0=deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0,
         quant_block_size=(quant_info.block_shape[1] if quant_info.block_shape else 128),
+        expert_alignment=block_e,
         expert_start=expert_start,
     )
     if packed_input_source is not hidden_states:
@@ -1080,6 +1156,7 @@ def pre_permute_standard_to_deep_gemm(
     running_state["hidden_states_device"] = hidden_states_device
     running_state["src2dst"] = src2dst
     running_state["all_tokens"] = all_tokens
+    running_state["contiguous_layout_alignment"] = block_e
     running_state["mxfp8_act_gran_k"] = (
         quant_info.block_shape[1] if quant_info.block_shape else 128
     )

python/sglang/srt/layers/quantization/fp8.py

@@ -2000,6 +2000,11 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 w2_q = layer.w2_weight.data
                 w13_s = layer.w13_weight_scale_inv.data
                 w2_s = layer.w2_weight_scale_inv.data
+            elif get_moe_runner_backend().is_cutlass():
+                w13_q = layer.w13_weight.data
+                w2_q = layer.w2_weight.data
+                w13_s = layer.w13_weight_scale_inv.data
+                w2_s = layer.w2_weight_scale_inv.data
             elif (
                 get_moe_runner_backend().is_flashinfer_trtllm()
                 or get_moe_runner_backend().is_flashinfer_trtllm_routed()
@@ -2554,6 +2559,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                 use_mxfp8=use_mxfp8,
                 output=symm_output,
                 enable_es=(use_mxfp8, use_mxfp8),
+                swiglu_limit=self.moe_runner_config.swiglu_limit,
             )
             return StandardCombineInput(hidden_states=output)
 

python/sglang/srt/layers/radix_attention.py

@@ -312,6 +312,7 @@ def _unified_attention_with_output_impl(
     q_rope: Optional[torch.Tensor] = None,
     k_rope: Optional[torch.Tensor] = None,
     sinks: Optional[torch.Tensor] = None,
+    attn_sink: Optional[torch.Tensor] = None,
     # MLA / TRT-LLM / NSA paths pass these through RadixAttention.forward(**kwargs);
     # they must appear in the schema when --enforce-piecewise-cuda-graph is on.
     cos_sin_cache: Optional[torch.Tensor] = None,
@@ -364,6 +365,8 @@ def _unified_attention_with_output_impl(
         kwargs["k_rope"] = k_rope[:key_value_num_tokens]
     if sinks is not None:
         kwargs["sinks"] = sinks
+    if attn_sink is not None:
+        kwargs["attn_sink"] = attn_sink
     if cos_sin_cache is not None:
         kwargs["cos_sin_cache"] = cos_sin_cache
     if is_neox is not None:
@@ -438,6 +441,7 @@ def unified_attention_with_output(
     q_rope: Optional[torch.Tensor] = None,
     k_rope: Optional[torch.Tensor] = None,
     sinks: Optional[torch.Tensor] = None,
+    attn_sink: Optional[torch.Tensor] = None,
     cos_sin_cache: Optional[torch.Tensor] = None,
     is_neox: Optional[bool] = None,
     llama_4_scaling: Optional[torch.Tensor] = None,
@@ -456,6 +460,7 @@ def unified_attention_with_output(
         q_rope=q_rope,
         k_rope=k_rope,
         sinks=sinks,
+        attn_sink=attn_sink,
         cos_sin_cache=cos_sin_cache,
         is_neox=is_neox,
         llama_4_scaling=llama_4_scaling,
@@ -486,6 +491,7 @@ def unified_attention_with_output_and_lse(
     q_rope: Optional[torch.Tensor] = None,
     k_rope: Optional[torch.Tensor] = None,
     sinks: Optional[torch.Tensor] = None,
+    attn_sink: Optional[torch.Tensor] = None,
     cos_sin_cache: Optional[torch.Tensor] = None,
     is_neox: Optional[bool] = None,
     llama_4_scaling: Optional[torch.Tensor] = None,
@@ -504,6 +510,7 @@ def unified_attention_with_output_and_lse(
         q_rope=q_rope,
         k_rope=k_rope,
         sinks=sinks,
+        attn_sink=attn_sink,
         cos_sin_cache=cos_sin_cache,
         is_neox=is_neox,
         llama_4_scaling=llama_4_scaling,

python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py

@@ -118,6 +118,18 @@ def should_defer_dsa_cp_kv_gather(
     return dsa_prefill_cp and fuse_rope_for_trtllm_mla
 
 
+def _apply_attention_output_gate(module, attn_output, gate):
+    apply_gate = getattr(module, "apply_attention_output_gate", None)
+    if apply_gate is not None:
+        return apply_gate(attn_output, gate)
+    if hasattr(module, "_apply_gated"):
+        return module._apply_gated(attn_output, gate)
+    raise RuntimeError(
+        "Prepared MLA attention gates are unsigmoided and require a "
+        "model-specific application hook"
+    )
+
+
 class DeepseekMLAForwardMixin:
     def init_mla_forward(self: DeepseekV2AttentionMLA):
         self.flashinfer_mla_disable_ragged = (
@@ -161,6 +173,8 @@ class DeepseekMLAForwardMixin:
             return False
         if is_kv_b_lora_active(self):
             return False
+        if getattr(self, "learnable_sink_param", None) is not None:
+            return False
         # The isolated 1-kernel graph is the bf16 fallback BMM. The fp8 and
         # DeepGEMM branches already use different fused paths.
         if self.w_kc.dtype == torch.float8_e4m3fn:
@@ -291,6 +305,11 @@ class DeepseekMLAForwardMixin:
         # True between the alt-stream fork and its consumption in the born
         # block; also suppresses the duplicate split/rope on that path.
         self._q8kv8_qprep_overlap_pending = False
+        attention_output_gate = (
+            self.prepare_attention_output_gate(hidden_states)
+            if hasattr(self, "prepare_attention_output_gate")
+            else None
+        )
 
         fuse_bmm_attention = (
             self.q_lora_rank is not None
@@ -667,6 +686,13 @@ class DeepseekMLAForwardMixin:
             topk_indices,
             llama_4_scaling,
             fusion_plan,
+            # Bailing's DsV3MLA appends its own gate to inner_state, so this
+            # slot is emitted only for models owning the gate hook.
+            *(
+                (attention_output_gate,)
+                if hasattr(self, "prepare_attention_output_gate")
+                else ()
+            ),
         )
 
     def forward_absorb_core(
@@ -681,18 +707,20 @@ class DeepseekMLAForwardMixin:
         topk_indices,
         llama_4_scaling,
         fusion_plan: Optional[MlaBmmFusionPlan] = None,
-        gate: Optional[torch.Tensor] = None,
+        attention_output_gate: Optional[torch.Tensor] = None,
     ):
         save_kv_cache = True
 
         if self.current_attention_backend in FORWARD_ABSORB_CORE_ATTENTION_BACKENDS:
             extra_args = {}
+            if getattr(self, "learnable_sink_param", None) is not None:
+                extra_args["attn_sink"] = self.learnable_sink_param
             if self._fuse_rope_for_trtllm_mla(forward_batch):
-                extra_args = {
-                    "cos_sin_cache": self.rotary_emb.cos_sin_cache,
-                    "is_neox": self.rotary_emb.is_neox_style,
-                    "llama_4_scaling": llama_4_scaling,
-                }
+                extra_args.update(
+                    cos_sin_cache=self.rotary_emb.cos_sin_cache,
+                    is_neox=self.rotary_emb.is_neox_style,
+                    llama_4_scaling=llama_4_scaling,
+                )
             if fusion_plan is not None:
                 bmm_attention_fn = (
                     bcg_mla_bmm_then_unified_attention
@@ -911,8 +939,10 @@ class DeepseekMLAForwardMixin:
             attn_bmm_output = apply_kv_b_lora_v_correction(
                 self, attn_output, attn_bmm_output
             )
-        if gate is not None:
-            attn_bmm_output = self._apply_gated(attn_bmm_output, gate)
+        if attention_output_gate is not None:
+            attn_bmm_output = _apply_attention_output_gate(
+                self, attn_bmm_output, attention_output_gate
+            )
         output, _ = self.o_proj(attn_bmm_output)
 
         if self.next_skip_topk is None:

python/sglang/srt/models/deepseek_common/deepseek_weight_loader.py

@@ -147,6 +147,44 @@ def _load_fused_indexer_wk(
     return True
 
 
+def _load_fused_expert_tensor(
+    name: str,
+    loaded_weight: torch.Tensor,
+    params_dict: Dict[str, torch.Tensor],
+) -> bool:
+    mappings = (
+        (".experts.gate_up_proj_scale", ".experts.w13_weight_scale_inv", "w13"),
+        (".experts.down_proj_scale", ".experts.w2_weight_scale_inv", "w2"),
+        (".experts.gate_up_proj", ".experts.w13_weight", "w13"),
+        (".experts.down_proj", ".experts.w2_weight", "w2"),
+    )
+    for source, target, shard_id in mappings:
+        if not name.endswith(source):
+            continue
+        param_name = name[: -len(source)] + target
+        param = params_dict.get(param_name)
+        if param is None:
+            return False
+        weight_loader = param.weight_loader
+        for expert_id, expert_weight in enumerate(loaded_weight):
+            if shard_id == "w13":
+                gate, up = expert_weight.chunk(2, dim=0)
+                weight_loader(
+                    param, gate, param_name, shard_id="w1", expert_id=expert_id
+                )
+                weight_loader(param, up, param_name, shard_id="w3", expert_id=expert_id)
+            else:
+                weight_loader(
+                    param,
+                    expert_weight,
+                    param_name,
+                    shard_id="w2",
+                    expert_id=expert_id,
+                )
+        return True
+    return False
+
+
 @dataclass(frozen=True)
 class NextNEnabledConfig:
     num_nextn_layers: int
@@ -281,6 +319,9 @@ class DeepseekV2WeightLoaderMixin:
                                 ):
                                     continue
 
+                if _load_fused_expert_tensor(name, loaded_weight, params_dict):
+                    continue
+
                 if "rotary_emb.inv_freq" in name:
                     continue
 
@@ -594,7 +635,11 @@ class DeepseekV2WeightLoaderMixin:
                         weight = w
 
                     # In multiple weight loading scenarios (e.g. RL), we need to inverse the scale of the weights after the requantization happened at the first loading.
-                    if (
+                    if weight_scale.format_ue8m0 and weight_scale.dtype == torch.uint8:
+                        weight_scale = (weight_scale.to(torch.int32) << 23).view(
+                            torch.float32
+                        )
+                    elif (
                         should_deepgemm_weight_requant_ue8m0(
                             weight_block_size=(
                                 self.quant_config.weight_block_size

python/sglang/srt/models/deepseek_v2.py

@@ -478,7 +478,14 @@ class MoEGate(nn.Module):
         self.is_nextn = is_nextn
         self.is_deepseek_v4 = is_deepseek_v4
         self.weight = nn.Parameter(
-            torch.empty((config.n_routed_experts, config.hidden_size))
+            torch.empty(
+                (config.n_routed_experts, config.hidden_size),
+                dtype=(
+                    torch.float32
+                    if getattr(config, "router_fp32", False)
+                    else torch.get_default_dtype()
+                ),
+            )
         )
 
         if config.topk_method == "noaux_tc" and not is_hash_moe:
@@ -515,6 +522,9 @@ class MoEGate(nn.Module):
         gemm_output_zero_allocator: BumpAllocator = None,
         forward_batch: ForwardBatch = None,
     ):
+        if self.weight.dtype == torch.float32:
+            return F.linear(hidden_states.float(), self.weight)
+
         if use_intel_amx_backend(self):
             return torch.ops.sgl_kernel.weight_packed_linear(
                 hidden_states,

python/sglang/srt/models/hunyuan_v4.py

@@ -0,0 +1,734 @@
+import functools
+import logging
+from typing import Iterable, Tuple
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from sglang.srt.distributed import get_pp_group
+from sglang.srt.layers.attention.index_topk_share import IndexTopKShareState
+from sglang.srt.layers.communicator import AttentionInputs, get_attn_tp_context
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import ColumnParallelLinear, ReplicatedLinear
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+    get_embedding_tp_kwargs,
+)
+from sglang.srt.model_executor.forward_context import get_attn_backend
+from sglang.srt.models.deepseek_common.attention_forward_methods import (
+    AttnForwardMethod,
+)
+from sglang.srt.models.deepseek_common.deepseek_weight_loader import (
+    DeepseekV2WeightLoaderMixin,
+)
+from sglang.srt.models.deepseek_v2 import (
+    DeepseekV2AttentionMLA,
+    DeepseekV2MLP,
+    DeepseekV2MoE,
+)
+from sglang.srt.runtime_context import get_parallel, get_stream
+from sglang.srt.utils import BumpAllocator, get_device_capability, is_cuda
+
+logger = logging.getLogger(__name__)
+
+
+def _hpc_ihc_available(op_name: str, hc_mult: int, hidden_size: int) -> bool:
+    try:
+        from sglang.kernels.ops.layernorm.hy4_ihc import _hpc_ihc_op
+    except ImportError:
+        return False
+    return _hpc_ihc_op(op_name, hc_mult, hidden_size) is not None
+
+
+def permute_hyv4_indexer_weight(name, loaded_weight, config):
+    if ".self_attn.indexer.wq_b." in name:
+        group_count = config.index_n_heads
+    elif any(
+        key in name
+        for key in (
+            ".self_attn.indexer.wk.",
+            ".self_attn.indexer.k_norm.",
+        )
+    ):
+        group_count = 1
+    else:
+        return loaded_weight
+
+    shape = loaded_weight.shape
+    loaded_weight = loaded_weight.reshape(
+        group_count,
+        config.index_head_dim,
+        *shape[1:],
+    )
+    rope_dim = config.qk_rope_head_dim
+    return torch.cat(
+        (
+            loaded_weight[:, -rope_dim:],
+            loaded_weight[:, :-rope_dim],
+        ),
+        dim=1,
+    ).reshape(shape)
+
+
+class HYV4HCPreLayer(nn.Module):
+    def __init__(self, config: PretrainedConfig, prefix: str):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.hc_mult = config.hc_mult
+        self.magnitude = config.hc_magnitude
+        self.hc_eps = config.hc_eps
+        self.rms_norm_eps = config.rms_norm_eps
+        self.hc_fn = ReplicatedLinear(
+            config.hidden_size * config.hc_mult,
+            2 * config.hc_mult,
+            bias=False,
+            params_dtype=torch.float32,
+            prefix=f"{prefix}.hc_fn",
+        )
+        self.hc_scale = nn.Parameter(torch.empty(2, dtype=torch.float32))
+        self.hc_base = nn.Parameter(
+            torch.empty(2 * config.hc_mult, dtype=torch.float32)
+        )
+        self._fused_ihc_pre_disabled = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        rms_weight: torch.Tensor | None = None,
+        rms_eps: float = 0.0,
+    ):
+        if hidden_states.is_cuda and not self._fused_ihc_pre_disabled:
+            try:
+                from sglang.kernels.ops.layernorm.hy4_ihc import fused_hy4_ihc_pre
+            except ImportError:
+                pass
+            else:
+                try:
+                    return fused_hy4_ihc_pre(
+                        hidden_states,
+                        self.hc_fn.weight,
+                        self.hc_scale,
+                        self.hc_base,
+                        self.magnitude,
+                        self.rms_norm_eps,
+                        self.hc_eps,
+                        rms_weight,
+                        rms_eps,
+                    )
+                except Exception:
+                    self._fused_ihc_pre_disabled = True
+                    logger.warning(
+                        "fused_hy4_ihc_pre failed, disabling fused path",
+                        exc_info=True,
+                    )
+        shape = hidden_states.shape
+        flat = hidden_states.flatten(1).float()
+        scale = torch.rsqrt(flat.square().mean(-1, keepdim=True) + self.rms_norm_eps)
+        gates = self.hc_fn(flat)[0] * scale
+        pre = (
+            torch.sigmoid(
+                gates[..., : self.hc_mult] * self.hc_scale[0]
+                + self.hc_base[: self.hc_mult]
+            )
+            + self.hc_eps
+        )
+        post = (
+            self.magnitude
+            * torch.sigmoid(
+                gates[..., self.hc_mult :] * self.hc_scale[1]
+                + self.hc_base[self.hc_mult :]
+            )
+            + self.hc_eps
+        )
+        reduced = torch.sum(pre.unsqueeze(-1) * hidden_states.reshape(shape), dim=1)
+        reduced = reduced.to(hidden_states.dtype)
+        if rms_weight is not None:
+            reduced_float = reduced.float()
+            reduced = (
+                reduced_float
+                * torch.rsqrt(
+                    reduced_float.square().mean(dim=-1, keepdim=True) + rms_eps
+                )
+                * rms_weight.float()
+            ).to(hidden_states.dtype)
+        return reduced, post
+
+
+class HYV4HCLayer(nn.Module):
+    def __init__(self, config: PretrainedConfig, prefix: str):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.hc_mult = config.hc_mult
+        self.hc_pre = HYV4HCPreLayer(config, f"{prefix}.hc_pre")
+        self._fused_ihc_post_disabled = False
+        self._fused_ihc_post_pre_disabled = False
+
+    def prepare_input(self, hidden_states: torch.Tensor):
+        if hidden_states.ndim == 3:
+            return hidden_states
+        if hidden_states.ndim != 2:
+            raise RuntimeError(
+                f"iHC expects a 2D or 3D tensor, got {hidden_states.shape}"
+            )
+        if hidden_states.shape[-1] == self.hidden_size:
+            return hidden_states.unsqueeze(1).repeat(1, self.hc_mult, 1)
+        if hidden_states.shape[-1] == self.hidden_size * self.hc_mult:
+            return hidden_states.reshape(-1, self.hc_mult, self.hidden_size)
+        raise RuntimeError(
+            "iHC input width must equal hidden_size or hc_mult * hidden_size"
+        )
+
+    def pre(self, hidden_states: torch.Tensor, norm: RMSNorm | None = None):
+        fuse_norm = (
+            norm is not None
+            and hidden_states.is_cuda
+            and _hpc_ihc_available("fuse_ihc_pre", self.hc_mult, self.hidden_size)
+        )
+        reduced, post = self.hc_pre(
+            hidden_states,
+            norm.weight if fuse_norm else None,
+            norm.variance_epsilon if fuse_norm else 0.0,
+        )
+        if norm is not None and not fuse_norm:
+            reduced = norm(reduced)
+        return reduced, post, hidden_states
+
+    def post(self, output, residual, post):
+        if output.is_cuda and not self._fused_ihc_post_disabled:
+            try:
+                from sglang.kernels.ops.layernorm.hy4_ihc import fused_hy4_ihc_post
+            except ImportError:
+                pass
+            else:
+                try:
+                    return fused_hy4_ihc_post(output, residual, post)
+                except Exception:
+                    self._fused_ihc_post_disabled = True
+                    logger.warning(
+                        "fused_hy4_ihc_post failed, disabling fused path",
+                        exc_info=True,
+                    )
+        result = post.float().unsqueeze(-1) * output.float().unsqueeze(1)
+        return (result + residual.float()).to(output.dtype)
+
+    def post_pre(self, output, residual, post, next_layer, norm):
+        if (
+            output.is_cuda
+            and not self._fused_ihc_post_pre_disabled
+            and _hpc_ihc_available("fuse_ihc_post_pre", self.hc_mult, self.hidden_size)
+        ):
+            try:
+                from sglang.kernels.ops.layernorm.hy4_ihc import (
+                    fused_hy4_ihc_post_pre,
+                )
+            except ImportError:
+                pass
+            else:
+                try:
+                    next_residual, reduced, next_post = fused_hy4_ihc_post_pre(
+                        output,
+                        residual,
+                        post,
+                        next_layer.hc_pre.hc_fn.weight,
+                        next_layer.hc_pre.hc_scale,
+                        next_layer.hc_pre.hc_base,
+                        next_layer.hc_pre.magnitude,
+                        next_layer.hc_pre.rms_norm_eps,
+                        next_layer.hc_pre.hc_eps,
+                        norm.weight,
+                        norm.variance_epsilon,
+                    )
+                    return reduced, next_post, next_residual
+                except Exception:
+                    self._fused_ihc_post_pre_disabled = True
+                    logger.warning(
+                        "fused_hy4_ihc_post_pre failed, disabling fused path",
+                        exc_info=True,
+                    )
+        next_residual = self.post(output, residual, post)
+        next_residual = next_layer.prepare_input(next_residual)
+        return next_layer.pre(next_residual, norm)
+
+
+class HYV4HCHeadLayer(nn.Module):
+    def __init__(self, config: PretrainedConfig, prefix: str):
+        super().__init__()
+        self.config = config
+        self.hc_head_fn = ReplicatedLinear(
+            config.hc_mult * config.hidden_size,
+            config.hc_mult,
+            bias=False,
+            params_dtype=torch.float32,
+            prefix=f"{prefix}.hc_head_fn",
+        )
+        self.hc_head_scale = nn.Parameter(torch.empty(1, dtype=torch.float32))
+        self.hc_head_base = nn.Parameter(
+            torch.empty(config.hc_mult, dtype=torch.float32)
+        )
+        self._fused_ihc_head_disabled = False
+
+    def forward(self, hidden_states: torch.Tensor, norm: RMSNorm | None = None):
+        if (
+            hidden_states.is_cuda
+            and not self._fused_ihc_head_disabled
+            and _hpc_ihc_available(
+                "fuse_ihc_head", self.config.hc_mult, self.config.hidden_size
+            )
+        ):
+            try:
+                from sglang.kernels.ops.layernorm.hy4_ihc import fused_hy4_ihc_head
+            except ImportError:
+                pass
+            else:
+                try:
+                    return fused_hy4_ihc_head(
+                        hidden_states,
+                        self.hc_head_fn.weight,
+                        self.hc_head_scale,
+                        self.hc_head_base,
+                        self.config.rms_norm_eps,
+                        self.config.hc_eps,
+                        None if norm is None else norm.weight,
+                        0.0 if norm is None else norm.variance_epsilon,
+                    )
+                except Exception:
+                    self._fused_ihc_head_disabled = True
+                    logger.warning(
+                        "fused_hy4_ihc_head failed, disabling fused path",
+                        exc_info=True,
+                    )
+        shape = hidden_states.shape
+        flat = hidden_states.flatten(1).float()
+        scale = torch.rsqrt(
+            flat.square().mean(-1, keepdim=True) + self.config.rms_norm_eps
+        )
+        gates = self.hc_head_fn(flat)[0] * scale
+        gates = (
+            torch.sigmoid(gates * self.hc_head_scale + self.hc_head_base)
+            + self.config.hc_eps
+        )
+        output = torch.sum(gates.unsqueeze(-1) * flat.reshape(shape), dim=1)
+        output = output.to(hidden_states.dtype)
+        return output if norm is None else norm(output)
+
+
+class HYV4Attention(DeepseekV2AttentionMLA):
+    def __init__(
+        self,
+        config,
+        layer_id,
+        quant_config=None,
+        prefix="",
+        alt_stream=None,
+        is_nextn=False,
+    ):
+        rope_parameters = config.rope_parameters
+        super().__init__(
+            config=config,
+            hidden_size=config.hidden_size,
+            num_heads=config.num_attention_heads,
+            qk_nope_head_dim=config.qk_nope_head_dim,
+            qk_rope_head_dim=config.qk_rope_head_dim,
+            v_head_dim=config.v_head_dim,
+            q_lora_rank=config.q_lora_rank,
+            kv_lora_rank=config.kv_lora_rank,
+            rope_theta=rope_parameters["rope_theta"],
+            rope_scaling=(
+                None
+                if rope_parameters.get("rope_type") == "default"
+                else rope_parameters
+            ),
+            max_position_embeddings=config.max_position_embeddings,
+            quant_config=quant_config,
+            reduce_results=True,
+            layer_id=layer_id,
+            prefix=prefix,
+            alt_stream=alt_stream,
+            is_nextn=is_nextn,
+        )
+        parallel = get_parallel()
+        self.linear_gate = ColumnParallelLinear(
+            config.hidden_size,
+            config.num_attention_heads * config.v_head_dim,
+            bias=False,
+            quant_config=quant_config,
+            tp_rank=parallel.attn_tp_rank,
+            tp_size=parallel.attn_tp_size,
+            prefix=f"{prefix}.linear_gate",
+        )
+        self.local_gate_width = self.num_local_heads * config.v_head_dim
+        if self.linear_gate.output_size_per_partition != self.local_gate_width:
+            raise ValueError(
+                "HYV4 attention gate shard width must match the local attention "
+                f"output width: {self.linear_gate.output_size_per_partition} != "
+                f"{self.local_gate_width}"
+            )
+        self.learnable_sink_param = nn.Parameter(
+            torch.empty(self.num_local_heads, dtype=torch.float32)
+        )
+        self.learnable_sink_param.weight_loader = self._sink_weight_loader
+        self._gate_fallback_backend = self._resolve_gate_fallback_backend()
+        self._gate_backend = self._resolve_gate_backend(
+            getattr(config, "gating_type", None), self._gate_fallback_backend
+        )
+        if prefix.endswith(".0.self_attn"):
+            logger.info("HYV4 MLA output gate backend: %s", self._gate_backend)
+
+    @staticmethod
+    def _sink_weight_loader(param, loaded_weight):
+        parallel = get_parallel()
+        heads = loaded_weight.shape[0] // parallel.attn_tp_size
+        start = parallel.attn_tp_rank * heads
+        param.data.copy_(loaded_weight[start : start + heads].float())
+
+    @staticmethod
+    def _resolve_gate_fallback_backend() -> str:
+        try:
+            from sglang.kernels.ops.moe.triton_sigmoid_gate_mul import (  # noqa: F401
+                sigmoid_gate_mul,
+            )
+        except ImportError:
+            return "eager"
+        return "triton"
+
+    def _resolve_gate_backend(
+        self, gating_type: str | None, fallback_backend: str
+    ) -> str:
+        weight = getattr(self.linear_gate, "weight", None)
+        if self._hpc_gated_mla_supported(
+            gating_type,
+            None if weight is None else weight.dtype,
+            None if weight is None else tuple(weight.shape),
+            self.local_gate_width,
+            self.hidden_size,
+        ):
+            return "hpc"
+        return fallback_backend
+
+    @staticmethod
+    @functools.lru_cache(maxsize=None)
+    def _hpc_gated_mla_supported(
+        gating_type: str | None,
+        weight_dtype: torch.dtype | None,
+        weight_shape: tuple[int, ...] | None,
+        local_gate_width: int,
+        hidden_size: int,
+    ) -> bool:
+        try:
+            import hpc
+        except ImportError:
+            return False
+
+        if not hasattr(getattr(hpc, "gemm", None), "gated_mla_gemm"):
+            logger.info(
+                "HY4 gated MLA: the installed hpc build (%s) has no "
+                "gemm.gated_mla_gemm; falling back.",
+                getattr(hpc, "__version__", "unknown"),
+            )
+            return False
+
+        major, minor = get_device_capability()
+        if (major, minor) not in ((10, 0), (10, 3)):
+            logger.warning(
+                "HY4 gated MLA: hpc.gated_mla_gemm is built for sm100/sm103, "
+                "got sm%d%d; falling back.",
+                major,
+                minor,
+            )
+            return False
+
+        # Headwise gating broadcasts one scalar per head over v_head_dim; the
+        # kernel only implements the elementwise product.
+        if gating_type != "elementwise":
+            return False
+
+        if weight_dtype != torch.bfloat16:
+            logger.warning(
+                "HY4 gated MLA: gate weight dtype is %s, the kernel needs "
+                "bfloat16 (keep linear_gate out of quantization).",
+                weight_dtype,
+            )
+            return False
+        expected_shape = (local_gate_width, hidden_size)
+        if weight_shape != expected_shape:
+            logger.warning(
+                "HYV4 gated MLA: local gate weight shape is %s, expected %s.",
+                weight_shape,
+                expected_shape,
+            )
+            return False
+        if local_gate_width % 256 != 0:
+            logger.warning(
+                "HY4 gated MLA: local gate width %s is not a multiple of 256.",
+                local_gate_width,
+            )
+            return False
+        return True
+
+    def _hpc_gated_mla_inputs_supported(self, hidden_states, attn_out) -> bool:
+        weight = self.linear_gate.weight
+        return (
+            hidden_states.dtype == torch.bfloat16
+            and weight.dtype == torch.bfloat16
+            and attn_out.dtype == torch.bfloat16
+            and hidden_states.ndim == 2
+            and weight.ndim == 2
+            and attn_out.ndim == 2
+            and hidden_states.shape[0] == attn_out.shape[0]
+            and hidden_states.shape[1] == weight.shape[1]
+            and weight.shape[0] == self.local_gate_width
+            and attn_out.shape[1] == self.local_gate_width
+            and hidden_states.device == weight.device == attn_out.device
+        )
+
+    @staticmethod
+    def _apply_attention_output_gate_fallback(attn_out, gate, backend):
+        if gate.shape != attn_out.shape:
+            raise ValueError(
+                "HYV4 projected attention gate shape must match the local attention "
+                f"output shape: {tuple(gate.shape)} != {tuple(attn_out.shape)}"
+            )
+        if backend == "triton" and attn_out.is_cuda:
+            from sglang.kernels.ops.moe.triton_sigmoid_gate_mul import (
+                sigmoid_gate_mul,
+            )
+
+            return sigmoid_gate_mul(attn_out, gate)
+        return attn_out * torch.sigmoid(gate)
+
+    def prepare_attention_output_gate(self, hidden_states):
+        # The hpc kernel fuses the projection in, so hand it the activations and
+        # let apply_attention_output_gate do the GEMM. The other two tiers want
+        # the projected gate.
+        if self._gate_backend == "hpc":
+            return hidden_states
+        return self.linear_gate(hidden_states)[0]
+
+    def apply_attention_output_gate(self, attn_out, gate):
+        if self._gate_backend == "hpc" and self._hpc_gated_mla_inputs_supported(
+            gate, attn_out
+        ):
+            from hpc.gemm import gated_mla_gemm
+
+            # linear_gate is column-parallel and unbiased, so the local weight
+            # shard lines up with the local attn_out columns.
+            return gated_mla_gemm(
+                gate.contiguous(), self.linear_gate.weight, attn_out.contiguous()
+            )
+        if self._gate_backend == "hpc":
+            gate = self.linear_gate(gate)[0]
+            backend = self._gate_fallback_backend
+        else:
+            backend = self._gate_backend
+        return self._apply_attention_output_gate_fallback(attn_out, gate, backend)
+
+    def dispatch_attn_forward_method(self, forward_batch):
+        backend = get_attn_backend()
+        backend = getattr(backend, "primary", backend)
+        if hasattr(backend, "use_mha") and backend.use_mha is not False:
+            backend.use_mha = False
+        method = super().dispatch_attn_forward_method(forward_batch)
+        if method != AttnForwardMethod.MLA:
+            raise RuntimeError("HYV4 requires the sparse MLA attention path")
+        return method
+
+
+class HYV4DecoderLayer(nn.Module):
+    def __init__(self, config, layer_id, quant_config=None, prefix="", alt_stream=None):
+        super().__init__()
+        self.self_attn = HYV4Attention(
+            config, layer_id, quant_config, f"{prefix}.self_attn", alt_stream
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        if config.mlp_layer_types[layer_id] == "dense":
+            self.mlp = DeepseekV2MLP(
+                config.hidden_size,
+                config.intermediate_size,
+                config.hidden_act,
+                quant_config=quant_config,
+                prefix=f"{prefix}.mlp",
+            )
+        else:
+            self.mlp = DeepseekV2MoE(
+                config,
+                layer_id,
+                quant_config=quant_config,
+                prefix=f"{prefix}.mlp",
+                alt_stream=alt_stream,
+            )
+            if hasattr(self.mlp, "shared_experts"):
+                self.mlp.shared_experts.swiglu_limit = None
+        self.hc_attn_layer = HYV4HCLayer(config, f"{prefix}.hc_attn_layer")
+        self.hc_mlp_layer = HYV4HCLayer(config, f"{prefix}.hc_mlp_layer")
+
+    def forward(
+        self,
+        positions,
+        hidden_states,
+        forward_batch,
+        zero_allocator,
+        prev_topk_indices=None,
+    ):
+        hidden_states = self.hc_attn_layer.prepare_input(hidden_states)
+        hidden_states, post, residual = self.hc_attn_layer.pre(
+            hidden_states, self.input_layernorm
+        )
+        get_attn_tp_context().set_attn_inputs(
+            AttentionInputs(
+                hidden_states, forward_batch, self.self_attn.prepare_qkv_latent
+            )
+        )
+        try:
+            hidden_states = self.self_attn(
+                positions,
+                hidden_states,
+                forward_batch,
+                zero_allocator,
+                prev_topk_indices=prev_topk_indices,
+            )
+        finally:
+            get_attn_tp_context().clear_attn_inputs()
+        if isinstance(hidden_states, tuple):
+            hidden_states, topk_indices = hidden_states
+        else:
+            topk_indices = None
+        hidden_states, post, residual = self.hc_attn_layer.post_pre(
+            hidden_states,
+            residual,
+            post,
+            self.hc_mlp_layer,
+            self.post_attention_layernorm,
+        )
+        if isinstance(self.mlp, DeepseekV2MoE):
+            hidden_states = self.mlp(hidden_states, forward_batch)
+        else:
+            hidden_states = self.mlp(hidden_states)
+        hidden_states = self.hc_mlp_layer.post(hidden_states, residual, post)
+        return hidden_states, topk_indices
+
+
+class HYV4Model(nn.Module):
+    def __init__(self, config, quant_config=None, prefix=""):
+        super().__init__()
+        if get_pp_group().world_size != 1:
+            raise ValueError("HYV4 pipeline parallelism is not supported")
+        self.config = config
+        self.start_layer = 0
+        self.end_layer = config.num_hidden_layers
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+            prefix=f"{prefix}.embed_tokens",
+            **get_embedding_tp_kwargs(),
+        )
+        self.alt_stream = get_stream("alt") if is_cuda() else None
+        self.layers = nn.ModuleList(
+            [
+                HYV4DecoderLayer(
+                    config,
+                    i,
+                    quant_config,
+                    f"{prefix}.layers.{i}",
+                    self.alt_stream,
+                )
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.hc_head = HYV4HCHeadLayer(config, f"{prefix}.hc_head")
+        self.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+
+    def forward(self, input_ids, positions, forward_batch, input_embeds=None):
+        hidden_states = (
+            self.embed_tokens(input_ids) if input_embeds is None else input_embeds
+        )
+        zero_allocator = BumpAllocator(
+            buffer_size=2 * len(self.layers),
+            dtype=torch.float32,
+            device=hidden_states.device,
+        )
+        topk_share = IndexTopKShareState(forward_batch, None)
+        for layer in self.layers:
+            hidden_states, topk_indices = layer(
+                positions,
+                hidden_states,
+                forward_batch,
+                zero_allocator,
+                topk_share.topk_indices,
+            )
+            topk_share.update(topk_indices)
+        topk_share.publish()
+        return self.hc_head(hidden_states, self.norm)
+
+
+class HYV4ForCausalLM(nn.Module, DeepseekV2WeightLoaderMixin):
+    packed_modules_mapping = {"gate_up_proj": ["gate_proj", "up_proj"]}
+
+    @classmethod
+    def shared_experts_fusion_disable_reason(cls, hf_config, quant_config):
+        return "HYV4 applies the SwiGLU limit only to routed experts."
+
+    def __init__(self, config, quant_config=None, prefix=""):
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.pp_group = get_pp_group()
+        self.model = HYV4Model(config, quant_config, f"{prefix}.model")
+        self.num_fused_shared_experts = max(
+            (
+                layer.mlp.num_fused_shared_experts
+                for layer in self.model.layers
+                if isinstance(layer.mlp, DeepseekV2MoE)
+            ),
+            default=0,
+        )
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            # Keep checkpoint weights in bf16; LogitsProcessor emits fp32
+            # logits when config.enable_lm_head_fp32 is set.
+            quant_config=quant_config,
+            prefix=f"{prefix}.lm_head",
+            use_attn_tp_group=get_parallel().enable_dp_lm_head,
+        )
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    def forward(self, input_ids, positions, forward_batch, input_embeds=None):
+        hidden_states = self.model(
+            input_ids, positions, forward_batch, input_embeds=input_embeds
+        )
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head, forward_batch
+        )
+
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_embed_and_head(self, embed, head):
+        del self.model.embed_tokens.weight
+        del self.lm_head.weight
+        self.model.embed_tokens.weight = embed
+        self.lm_head.weight = head
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        def mapped_weights():
+            for name, loaded_weight in weights:
+                if name.startswith("model.mtp_layers."):
+                    continue
+                loaded_weight = permute_hyv4_indexer_weight(
+                    name, loaded_weight, self.config
+                )
+                if name.endswith((".hc_fn", ".hc_head_fn")):
+                    name += ".weight"
+                if name.endswith(".weight_scale"):
+                    name += "_inv"
+                yield name, loaded_weight
+
+        self.do_load_weights(mapped_weights())
+
+
+EntryClass = [HYV4ForCausalLM]

python/sglang/srt/models/hunyuan_v4_nextn.py

@@ -0,0 +1,229 @@
+import copy
+from typing import Iterable, Tuple
+
+import torch
+from torch import nn
+
+from sglang.srt.distributed import get_pp_group
+from sglang.srt.layers.attention.index_topk_share import IndexTopKShareState
+from sglang.srt.layers.communicator import AttentionInputs, get_attn_tp_context
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+    get_embedding_tp_kwargs,
+)
+from sglang.srt.models.deepseek_common.deepseek_weight_loader import (
+    DeepseekV2WeightLoaderMixin,
+)
+from sglang.srt.models.deepseek_v2 import DeepseekV2MoE
+from sglang.srt.models.hunyuan_v4 import (
+    HYV4Attention,
+    HYV4ForCausalLM,
+    permute_hyv4_indexer_weight,
+)
+from sglang.srt.runtime_context import get_parallel, get_stream
+from sglang.srt.utils import BumpAllocator, is_cuda
+
+
+def _mtp_quant_config(quant_config):
+    if quant_config is None:
+        return None
+    quant_config = copy.deepcopy(quant_config)
+    ignored_layers = getattr(quant_config, "ignored_layers", None)
+    if ignored_layers is not None:
+        quant_config.ignored_layers = [
+            name.replace("model.mtp_layers.0", "model.decoder").replace(
+                "mtp_layers.0", "model.decoder"
+            )
+            for name in ignored_layers
+        ]
+    return quant_config
+
+
+class HYV4MTPDecoderLayer(nn.Module):
+    def __init__(self, config, quant_config=None, prefix="", alt_stream=None):
+        super().__init__()
+        self.self_attn = HYV4Attention(
+            config,
+            0,
+            quant_config,
+            f"{prefix}.self_attn",
+            alt_stream,
+            is_nextn=True,
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.mlp = DeepseekV2MoE(
+            config,
+            0,
+            quant_config=quant_config,
+            prefix=f"{prefix}.mlp",
+            alt_stream=alt_stream,
+            is_nextn=True,
+        )
+        if hasattr(self.mlp, "shared_experts"):
+            self.mlp.shared_experts.swiglu_limit = None
+
+    def forward(
+        self,
+        positions,
+        hidden_states,
+        forward_batch,
+        zero_allocator,
+        prev_topk_indices=None,
+    ):
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        get_attn_tp_context().set_attn_inputs(
+            AttentionInputs(
+                hidden_states, forward_batch, self.self_attn.prepare_qkv_latent
+            )
+        )
+        try:
+            hidden_states = self.self_attn(
+                positions,
+                hidden_states,
+                forward_batch,
+                zero_allocator,
+                prev_topk_indices=prev_topk_indices,
+            )
+        finally:
+            get_attn_tp_context().clear_attn_inputs()
+        if isinstance(hidden_states, tuple):
+            hidden_states, topk_indices = hidden_states
+        else:
+            topk_indices = None
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states, forward_batch)
+        return hidden_states, residual, topk_indices
+
+
+class HYV4ModelNextN(nn.Module):
+    def __init__(self, config, quant_config=None, prefix=""):
+        super().__init__()
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+            prefix=f"{prefix}.embed_tokens",
+            **get_embedding_tp_kwargs(),
+        )
+        self.enorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.hnorm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+        self.eh_proj = nn.Linear(2 * config.hidden_size, config.hidden_size, bias=False)
+        self.alt_stream = get_stream("alt") if is_cuda() else None
+        self.decoder = HYV4MTPDecoderLayer(
+            config,
+            quant_config,
+            f"{prefix}.decoder",
+            self.alt_stream,
+        )
+        self.shared_head = nn.Module()
+        self.shared_head.norm = RMSNorm(config.hidden_size, config.rms_norm_eps)
+
+    def forward(self, input_ids, positions, forward_batch, input_embeds=None):
+        hidden_states = (
+            self.embed_tokens(input_ids) if input_embeds is None else input_embeds
+        )
+        if hidden_states.shape[0] > 0:
+            hidden_states = self.eh_proj(
+                torch.cat(
+                    (
+                        self.enorm(hidden_states),
+                        self.hnorm(forward_batch.spec_info.hidden_states),
+                    ),
+                    dim=-1,
+                )
+            )
+        zero_allocator = BumpAllocator(
+            buffer_size=2,
+            dtype=torch.float32,
+            device=hidden_states.device,
+        )
+        topk_share = IndexTopKShareState.from_mtp_carry(forward_batch)
+        hidden_states, residual, topk_indices = self.decoder(
+            positions,
+            hidden_states,
+            forward_batch,
+            zero_allocator,
+            topk_share.topk_indices,
+        )
+        topk_share.update(topk_indices)
+        topk_share.publish()
+        if forward_batch.forward_mode.is_idle():
+            return hidden_states
+        hidden_states, _ = self.shared_head.norm(hidden_states, residual)
+        return hidden_states
+
+
+class HYV4ForCausalLMNextN(nn.Module, DeepseekV2WeightLoaderMixin):
+    packed_modules_mapping = {"gate_up_proj": ["gate_proj", "up_proj"]}
+
+    @staticmethod
+    def shared_experts_fusion_disable_reason(hf_config, quant_config):
+        return HYV4ForCausalLM.shared_experts_fusion_disable_reason(
+            hf_config, quant_config
+        )
+
+    def __init__(self, config, quant_config=None, prefix=""):
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.pp_group = get_pp_group()
+        nextn_quant_config = _mtp_quant_config(quant_config)
+        self.model = HYV4ModelNextN(
+            config, nextn_quant_config, prefix=f"{prefix}.model"
+        )
+        self.num_fused_shared_experts = self.model.decoder.mlp.num_fused_shared_experts
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+            prefix=f"{prefix}.lm_head",
+            use_attn_tp_group=get_parallel().enable_dp_lm_head,
+        )
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    def forward(self, input_ids, positions, forward_batch):
+        hidden_states = self.model(input_ids, positions, forward_batch)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head, forward_batch
+        )
+
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_embed_and_head(self, embed, head):
+        del self.model.embed_tokens.weight
+        del self.lm_head.weight
+        self.model.embed_tokens.weight = embed
+        self.lm_head.weight = head
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        layer_prefix = f"model.layers.{self.config.num_hidden_layers}"
+
+        def mapped_weights():
+            for name, loaded_weight in weights:
+                if not name.startswith("model.mtp_layers.0."):
+                    continue
+                name = name.replace("model.mtp_layers.0", layer_prefix)
+                if name.endswith(".final_layernorm.weight"):
+                    name = name.replace(
+                        ".final_layernorm.weight", ".shared_head.norm.weight"
+                    )
+                loaded_weight = permute_hyv4_indexer_weight(
+                    name, loaded_weight, self.config
+                )
+                if name.endswith(".weight_scale"):
+                    name += "_inv"
+                yield name, loaded_weight
+
+        self.do_load_weights(mapped_weights(), is_nextn=True)
+
+    def post_load_weights(self, is_nextn=True, weight_names=None):
+        super().post_load_weights(is_nextn=True, weight_names=weight_names)
+
+
+EntryClass = [HYV4ForCausalLMNextN]

python/sglang/srt/parser/hunyuan_reasoning.py

@@ -0,0 +1,44 @@
+from typing import Optional
+
+from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
+from sglang.srt.parser.template_detection import ReasoningToggleConfig
+
+
+def uses_hunyuan_reasoning_effort(
+    reasoning_parser: Optional[str], reasoning_config: Optional[ReasoningToggleConfig]
+) -> bool:
+    return (
+        reasoning_parser == "hunyuan"
+        and reasoning_config is not None
+        and reasoning_config.special_case == "hunyuan_effort"
+    )
+
+
+def normalize_hunyuan_reasoning_effort(
+    request: ChatCompletionRequest,
+    reasoning_parser: Optional[str],
+    reasoning_config: Optional[ReasoningToggleConfig],
+) -> None:
+    if not uses_hunyuan_reasoning_effort(reasoning_parser, reasoning_config):
+        return
+
+    effort = request.reasoning_effort
+    if effort is None and request.chat_template_kwargs is not None:
+        effort = request.chat_template_kwargs.get("reasoning_effort")
+    if effort is None:
+        normalized_effort = "high"
+    elif effort in ("none", "no_think"):
+        normalized_effort = "no_think"
+    elif effort in ("minimal", "low"):
+        normalized_effort = "low"
+    elif effort in ("medium", "high", "xhigh", "max"):
+        normalized_effort = "high"
+    else:
+        raise ValueError(
+            "Hunyuan reasoning_effort must be one of none, minimal, low, "
+            "medium, high, xhigh, or max"
+        )
+
+    request.reasoning_effort = normalized_effort
+    if request.chat_template_kwargs is not None:
+        request.chat_template_kwargs.pop("reasoning_effort", None)

python/sglang/srt/parser/template_detection.py

@@ -175,6 +175,15 @@ REASONING_MODE_RULES = (
             ctx.has_text("reasoning_effort") and ctx.has_text("[THINK]")
         ),
     ),
+    DetectionRule(
+        name="hunyuan_reasoning_effort",
+        value=ReasoningToggleConfig(special_case="hunyuan_effort"),
+        predicate=lambda ctx: (
+            ctx.has_text("reasoning_effort")
+            and ctx.has_text("reasoning_mode_token")
+            and ctx.has_text("no_think")
+        ),
+    ),
     DetectionRule(
         name="explicit_enable_thinking_default_false",
         value=ReasoningToggleConfig(
@@ -364,8 +373,16 @@ def _is_hunyuan(ctx):
     sep = ctx.has_text("<tool_sep>") or ctx.has_vocab_pattern(
         r"^<tool_sep(?::[^>]+)?>$"
     )
-    return (tc and sep) or (
-        ctx.has_text("reasoning_effort") and ctx.has_text("interleaved_thinking")
+    return (
+        (tc and sep)
+        or (
+            tc
+            and ctx.reasoning_config
+            == ReasoningToggleConfig(special_case="hunyuan_effort")
+            and ctx.has_vocab_pattern(r"^<arg_key(?::[^>]+)?>$")
+            and ctx.has_vocab_pattern(r"^<arg_value(?::[^>]+)?>$")
+        )
+        or (ctx.has_text("reasoning_effort") and ctx.has_text("interleaved_thinking"))
     )
 
 

python/sglang/srt/utils/hf_transformers/common.py

@@ -38,6 +38,7 @@ from sglang.srt.configs import (
     ExaoneConfig,
     FalconH1Config,
     GraniteMoeHybridConfig,
+    HYV4Config,
     InklingAudioConfig,
     InklingMMConfig,
     InklingModelConfig,
@@ -122,6 +123,7 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
         Qwen3NextConfig,
         FalconH1Config,
         GraniteMoeHybridConfig,
+        HYV4Config,
         DotsVLMConfig,
         DotsOCRConfig,
         Dots3Config,

test/registered/kernels/ops/attention/test_dsa_sink_pad_cache.py

@@ -0,0 +1,47 @@
+import sys
+
+import pytest
+import sgl_kernel.flash_mla as flash_mla
+import torch
+
+from sglang.srt.layers.attention.dsa_backend import DeepseekSparseAttnBackend
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=5, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA is required")
+def test_sparse_flashmla_sink_padding_refreshes_reused_buffer(monkeypatch):
+    backend = object.__new__(DeepseekSparseAttnBackend)
+    backend.device_sm_major = 10
+    backend.dsa_index_topk = 2
+    backend._sink_pad_cache = {}
+    captured_sinks = []
+
+    def capture_flash_mla_sparse_fwd(**kwargs):
+        captured_sinks.append(kwargs["attn_sink"].clone())
+        q = kwargs["q"]
+        return q.new_zeros((*q.shape[:2], kwargs["d_v"])), None, None
+
+    monkeypatch.setattr(flash_mla, "flash_mla_sparse_fwd", capture_flash_mla_sparse_fwd)
+
+    q = torch.zeros((1, 64, 8), device="cuda")
+    kv_cache = torch.zeros((1, 1, 8), device="cuda")
+    page_table = torch.zeros((1, 2), dtype=torch.int32, device="cuda")
+    sink = torch.arange(64, dtype=torch.float32, device="cuda")
+
+    backend._forward_flashmla_sparse(q, kv_cache, 8, page_table, 1.0, attn_sink=sink)
+    cached_sink = next(iter(backend._sink_pad_cache.values()))
+    cached_ptr = cached_sink.data_ptr()
+
+    sink.add_(100)
+    backend._forward_flashmla_sparse(q, kv_cache, 8, page_table, 1.0, attn_sink=sink)
+
+    assert next(iter(backend._sink_pad_cache.values())).data_ptr() == cached_ptr
+    torch.testing.assert_close(captured_sinks[0][:64], sink - 100)
+    torch.testing.assert_close(captured_sinks[1][:64], sink)
+    torch.testing.assert_close(captured_sinks[1][64:], torch.zeros_like(sink))
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__]))

test/registered/kernels/ops/attention/test_hy4_hpc_gated_mla.py

@@ -0,0 +1,145 @@
+import sys
+
+import pytest
+import torch
+from torch import nn
+
+from sglang.kernels.jit.utils import get_ci_test_range
+from sglang.srt.models.hunyuan_v4 import HYV4Attention
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(
+    est_time=60,
+    nightly=False,
+    disabled=None,
+    stage="base-b-kernel-unit",
+    runner_config="4-gpu-b200",
+)
+register_cuda_ci(
+    est_time=120,
+    nightly=True,
+    disabled=None,
+    stage="nightly",
+    runner_config="4-gpu-b200",
+)
+
+# Guards attention-TP shards from silently reverting to the global N=16384 shape.
+LOCAL_GATE_WIDTHS = get_ci_test_range(
+    full_range=[256, 512, 1024, 2048, 4096, 8192, 16384],
+    ci_range=[256, 2048, 16384],
+)
+
+
+def _hpc_gated_mla_available():
+    try:
+        from hpc.gemm import gated_mla_gemm  # noqa: F401
+    except (AttributeError, ImportError):
+        return False
+    return torch.cuda.is_available() and torch.cuda.get_device_capability() in (
+        (10, 0),
+        (10, 3),
+    )
+
+
+class TupleLinear(nn.Module):
+    def __init__(self, weight):
+        super().__init__()
+        self.weight = nn.Parameter(weight, requires_grad=False)
+
+    def forward(self, inputs):
+        return nn.functional.linear(inputs, self.weight), None
+
+
+def _make_attention(weight):
+    attention = HYV4Attention.__new__(HYV4Attention)
+    nn.Module.__init__(attention)
+    attention.linear_gate = TupleLinear(weight)
+    attention.local_gate_width = weight.shape[0]
+    attention._gate_backend = "hpc"
+    attention._gate_fallback_backend = "eager"
+    return attention
+
+
+@pytest.mark.skipif(
+    not _hpc_gated_mla_available(),
+    reason="requires HPC-Ops gated MLA on SM100 or SM103",
+)
+@pytest.mark.parametrize("local_gate_width", LOCAL_GATE_WIDTHS)
+def test_hy4_gated_mla_attention_tp_eager_graph_parity(local_gate_width):
+    torch.manual_seed(local_gate_width)
+    hidden_size = 6144
+    batch_size = 7
+    weight = torch.randn(
+        local_gate_width, hidden_size, dtype=torch.bfloat16, device="cuda"
+    )
+    hidden_states = torch.randn(
+        batch_size, hidden_size, dtype=torch.bfloat16, device="cuda"
+    )
+    attn_out = torch.randn(
+        batch_size, local_gate_width, dtype=torch.bfloat16, device="cuda"
+    )
+    attention = _make_attention(weight)
+
+    def run():
+        gate = attention.prepare_attention_output_gate(hidden_states)
+        return attention.apply_attention_output_gate(attn_out, gate)
+
+    expected = attn_out * torch.sigmoid(nn.functional.linear(hidden_states, weight))
+    eager = run()
+    torch.testing.assert_close(eager, expected, rtol=0.08, atol=0.01)
+
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        run()
+    torch.cuda.current_stream().wait_stream(stream)
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        graph_out = run()
+
+    graph.replay()
+    torch.testing.assert_close(graph_out, eager, rtol=0, atol=0)
+
+
+@pytest.mark.skipif(
+    not _hpc_gated_mla_available(),
+    reason="requires HPC-Ops gated MLA on SM100 or SM103",
+)
+@pytest.mark.parametrize("batch_size", [128, 129])
+def test_hy4_gated_mla_dispatch_boundary_graph_parity(batch_size):
+    torch.manual_seed(batch_size)
+    hidden_size = 6144
+    local_gate_width = 256
+    weight = torch.randn(
+        local_gate_width, hidden_size, dtype=torch.bfloat16, device="cuda"
+    )
+    hidden_states = torch.randn(
+        batch_size, hidden_size, dtype=torch.bfloat16, device="cuda"
+    )
+    attn_out = torch.randn(
+        batch_size, local_gate_width, dtype=torch.bfloat16, device="cuda"
+    )
+    attention = _make_attention(weight)
+
+    def run():
+        gate = attention.prepare_attention_output_gate(hidden_states)
+        return attention.apply_attention_output_gate(attn_out, gate)
+
+    eager = run()
+    expected = attn_out * torch.sigmoid(nn.functional.linear(hidden_states, weight))
+    torch.testing.assert_close(eager, expected, rtol=0.08, atol=0.01)
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        run()
+    torch.cuda.current_stream().wait_stream(stream)
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        graph_out = run()
+
+    graph.replay()
+    torch.testing.assert_close(graph_out, eager, rtol=0, atol=0)
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__]))

test/registered/kernels/ops/layernorm/test_hy4_decode_kernels.py

@@ -0,0 +1,229 @@
+import unittest
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import torch
+import triton
+import triton.language as tl
+from torch import nn
+
+import sglang.kernels.ops.layernorm.hy4_ihc as hy4_ihc
+from sglang.srt.models import hunyuan_v4
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cuda_ci(est_time=35, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+
+
+@triton.jit
+def _reference_hy4_ihc_pre_kernel(
+    x_ptr,
+    fn_ptr,
+    scale_ptr,
+    base_ptr,
+    y_ptr,
+    post_ptr,
+    hidden_size: tl.constexpr,
+    HC_MULT: tl.constexpr,
+    HC_POW2: tl.constexpr,
+    K_TOTAL: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    BLOCK_D: tl.constexpr,
+    magnitude: tl.constexpr,
+    norm_eps: tl.constexpr,
+    hc_eps: tl.constexpr,
+):
+    pid = tl.program_id(0).to(tl.int64)
+    x_row = x_ptr + pid * K_TOTAL
+    m_idx = tl.arange(0, HC_POW2)
+    m_mask = m_idx < HC_MULT
+
+    sumsq = tl.zeros((), dtype=tl.float32)
+    mix_pre = tl.zeros((HC_POW2,), dtype=tl.float32)
+    mix_post = tl.zeros((HC_POW2,), dtype=tl.float32)
+    for k_off in tl.range(0, K_TOTAL, BLOCK_K):
+        k_offs = k_off + tl.arange(0, BLOCK_K)
+        k_mask = k_offs < K_TOTAL
+        x_tile = tl.load(x_row + k_offs, mask=k_mask, other=0.0).to(tl.float32)
+        sumsq += tl.sum(x_tile * x_tile, axis=0)
+
+        fn_offs = m_idx[:, None] * K_TOTAL + k_offs[None, :]
+        fn_mask = m_mask[:, None] & k_mask[None, :]
+        mix_pre += tl.sum(
+            tl.load(fn_ptr + fn_offs, mask=fn_mask, other=0.0) * x_tile[None, :],
+            axis=1,
+        )
+        mix_post += tl.sum(
+            tl.load(
+                fn_ptr + HC_MULT * K_TOTAL + fn_offs,
+                mask=fn_mask,
+                other=0.0,
+            )
+            * x_tile[None, :],
+            axis=1,
+        )
+
+    rsqrt = tl.rsqrt(sumsq / K_TOTAL + norm_eps)
+    scale_pre = tl.load(scale_ptr)
+    scale_post = tl.load(scale_ptr + 1)
+    base_pre = tl.load(base_ptr + m_idx, mask=m_mask, other=0.0)
+    base_post = tl.load(base_ptr + HC_MULT + m_idx, mask=m_mask, other=0.0)
+
+    pre = tl.sigmoid(mix_pre * rsqrt * scale_pre + base_pre) + hc_eps
+    post = magnitude * tl.sigmoid(mix_post * rsqrt * scale_post + base_post) + hc_eps
+    tl.store(post_ptr + pid * HC_MULT + m_idx, post, mask=m_mask)
+
+    y_row = y_ptr + pid * hidden_size
+    for d_off in tl.range(0, hidden_size, BLOCK_D):
+        d_offs = d_off + tl.arange(0, BLOCK_D)
+        d_mask = d_offs < hidden_size
+        y_block = tl.zeros((BLOCK_D,), dtype=tl.float32)
+        for m in tl.static_range(HC_MULT):
+            x_m = tl.load(x_row + m * hidden_size + d_offs, mask=d_mask, other=0.0)
+            pre_m = tl.sum(tl.where(m_idx == m, pre, 0.0), axis=0)
+            y_block += pre_m * x_m.to(tl.float32)
+        tl.store(
+            y_row + d_offs,
+            y_block.to(y_ptr.dtype.element_ty),
+            mask=d_mask,
+        )
+
+
+def _reference_hy4_ihc_pre(x, hc_fn, hc_scale, hc_base):
+    num_tokens, hc_mult, hidden_size = x.shape
+    k_total = hc_mult * hidden_size
+    y = torch.empty((num_tokens, hidden_size), dtype=x.dtype, device=x.device)
+    post = torch.empty((num_tokens, hc_mult), dtype=torch.float32, device=x.device)
+    if num_tokens == 0:
+        return y, post
+
+    _reference_hy4_ihc_pre_kernel[(num_tokens,)](
+        x,
+        hc_fn,
+        hc_scale,
+        hc_base,
+        y,
+        post,
+        hidden_size=hidden_size,
+        HC_MULT=hc_mult,
+        HC_POW2=triton.next_power_of_2(hc_mult),
+        K_TOTAL=k_total,
+        BLOCK_K=1024,
+        BLOCK_D=1024,
+        magnitude=2.0,
+        norm_eps=1e-6,
+        hc_eps=1e-6,
+        num_warps=8,
+        enable_fp_fusion=False,
+    )
+    return y, post
+
+
+@unittest.skipUnless(torch.cuda.is_available(), "HYV4 Triton kernels need CUDA")
+class TestHy4DecodeKernels(CustomTestCase):
+    def test_split_k_matches_single_cta(self):
+        torch.manual_seed(0)
+        for num_tokens, hidden_size in (
+            (0, 6144),
+            (1, 4096),
+            (3, 4100),
+            (31, 6144),
+            (64, 6144),
+        ):
+            with self.subTest(num_tokens=num_tokens, hidden_size=hidden_size):
+                hc_mult = 4
+                k_total = hc_mult * hidden_size
+                x = torch.randn(
+                    num_tokens,
+                    hc_mult,
+                    hidden_size,
+                    device="cuda",
+                    dtype=torch.bfloat16,
+                )
+                hc_fn = (
+                    torch.randn(
+                        2 * hc_mult,
+                        k_total,
+                        device="cuda",
+                        dtype=torch.float32,
+                    )
+                    * 0.02
+                )
+                hc_scale = torch.tensor([0.7, 1.3], device="cuda", dtype=torch.float32)
+                hc_base = (
+                    torch.randn(2 * hc_mult, device="cuda", dtype=torch.float32) * 0.5
+                )
+
+                expected = _reference_hy4_ihc_pre(x, hc_fn, hc_scale, hc_base)
+                with patch.object(hy4_ihc, "_hpc_ihc_op", return_value=None):
+                    actual = hy4_ihc.fused_hy4_ihc_pre(
+                        x, hc_fn, hc_scale, hc_base, 2.0, 1e-6, 1e-6
+                    )
+
+                self.assertTrue(torch.equal(actual[0], expected[0]))
+                self.assertTrue(torch.equal(actual[1], expected[1]))
+
+    def test_fused_ihc_failure_disables_each_path(self):
+        class TupleLinear(nn.Module):
+            def __init__(self, input_size, output_size):
+                super().__init__()
+                self.weight = nn.Parameter(torch.randn(output_size, input_size))
+
+            def forward(self, inputs):
+                return nn.functional.linear(inputs, self.weight), None
+
+        config = SimpleNamespace(
+            hidden_size=8,
+            hc_mult=2,
+            hc_magnitude=2.0,
+            hc_eps=1e-6,
+            rms_norm_eps=1e-5,
+        )
+        counts = {"pre": 0, "post": 0, "post_pre": 0, "head": 0}
+
+        def fail(name):
+            def raise_error(*args, **kwargs):
+                counts[name] += 1
+                raise RuntimeError(name)
+
+            return raise_error
+
+        def make_linear(input_size, output_size, **kwargs):
+            return TupleLinear(input_size, output_size)
+
+        with patch.object(hunyuan_v4, "ReplicatedLinear", make_linear):
+            pre_layer = hunyuan_v4.HYV4HCPreLayer(config, "pre").cuda()
+            layer = hunyuan_v4.HYV4HCLayer(config, "layer").cuda()
+            next_layer = hunyuan_v4.HYV4HCLayer(config, "next").cuda()
+            head_layer = hunyuan_v4.HYV4HCHeadLayer(config, "head").cuda()
+
+        next_layer.hc_pre._fused_ihc_pre_disabled = True
+        norm = hunyuan_v4.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps, force_native=True
+        ).cuda()
+        hidden_states = torch.randn(3, 2, 8, device="cuda")
+        output = torch.randn(3, 8, device="cuda")
+        residual = torch.randn(3, 2, 8, device="cuda")
+        post = torch.randn(3, 2, device="cuda")
+
+        with (
+            patch.object(hy4_ihc, "fused_hy4_ihc_pre", fail("pre")),
+            patch.object(hy4_ihc, "fused_hy4_ihc_post", fail("post")),
+            patch.object(hy4_ihc, "fused_hy4_ihc_post_pre", fail("post_pre")),
+            patch.object(hy4_ihc, "fused_hy4_ihc_head", fail("head")),
+            patch.object(hunyuan_v4, "_hpc_ihc_available", return_value=True),
+            self.assertLogs(hunyuan_v4.logger, level="WARNING") as logs,
+        ):
+            for _ in range(2):
+                pre_layer(hidden_states)
+                layer.post(output, residual, post)
+                layer.post_pre(output, residual, post, next_layer, norm)
+                head_layer(hidden_states)
+
+        self.assertEqual(counts, {"pre": 1, "post": 1, "post_pre": 1, "head": 1})
+        self.assertEqual(len(logs.records), 4)
+        self.assertTrue(all(record.exc_info is not None for record in logs.records))
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/kernels/ops/layernorm/test_hy4_hpc_ihc.py

@@ -0,0 +1,111 @@
+import sys
+
+import pytest
+import torch
+
+from sglang.kernels.ops.layernorm.hy4_ihc import (
+    _hpc_ihc_op,
+    fused_hy4_ihc_head,
+    fused_hy4_ihc_post_pre,
+    fused_hy4_ihc_pre,
+)
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+
+
+@pytest.mark.parametrize("hidden_size", [4096, 6144])
+def test_hpc_ihc_eager_graph_parity(hidden_size):
+    if not torch.cuda.is_available():
+        pytest.skip("requires CUDA")
+    if _hpc_ihc_op("fuse_ihc_post_pre", 4, hidden_size) is None:
+        pytest.skip("requires a compatible HPC-Ops iHC build")
+
+    torch.manual_seed(13)
+    num_tokens = 7
+    hc_mult = 4
+    norm_eps, hc_eps, magnitude = 1e-5, 1e-6, 2.0
+    x = torch.rand(
+        (num_tokens, hc_mult, hidden_size), dtype=torch.bfloat16, device="cuda"
+    )
+    output = torch.rand((num_tokens, hidden_size), dtype=torch.bfloat16, device="cuda")
+    pre_weight = (
+        torch.rand(
+            (2 * hc_mult, hc_mult * hidden_size),
+            dtype=torch.float32,
+            device="cuda",
+        )
+        * 6e-3
+    )
+    next_weight = torch.rand_like(pre_weight) * 6e-3
+    head_weight = (
+        torch.rand(
+            (hc_mult, hc_mult * hidden_size),
+            dtype=torch.float32,
+            device="cuda",
+        )
+        * 6e-3
+    )
+    pre_scale = torch.rand((2,), dtype=torch.float32, device="cuda")
+    next_scale = torch.rand((2,), dtype=torch.float32, device="cuda")
+    head_scale = torch.rand((1,), dtype=torch.float32, device="cuda")
+    pre_base = torch.rand((2 * hc_mult,), dtype=torch.float32, device="cuda")
+    next_base = torch.rand((2 * hc_mult,), dtype=torch.float32, device="cuda")
+    head_base = torch.rand((hc_mult,), dtype=torch.float32, device="cuda")
+    rms_weight = torch.rand((hidden_size,), dtype=torch.bfloat16, device="cuda")
+
+    def run():
+        _, post = fused_hy4_ihc_pre(
+            x,
+            pre_weight,
+            pre_scale,
+            pre_base,
+            magnitude,
+            norm_eps,
+            hc_eps,
+            rms_weight,
+            norm_eps,
+        )
+        residual, reduced, next_post = fused_hy4_ihc_post_pre(
+            output,
+            x,
+            post,
+            next_weight,
+            next_scale,
+            next_base,
+            magnitude,
+            norm_eps,
+            hc_eps,
+            rms_weight,
+            norm_eps,
+        )
+        head = fused_hy4_ihc_head(
+            residual,
+            head_weight,
+            head_scale,
+            head_base,
+            norm_eps,
+            hc_eps,
+            rms_weight,
+            norm_eps,
+        )
+        return residual, reduced, next_post, head
+
+    warmup_stream = torch.cuda.Stream()
+    warmup_stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(warmup_stream):
+        run()
+    torch.cuda.current_stream().wait_stream(warmup_stream)
+
+    eager = tuple(value.clone() for value in run())
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        graph_outputs = run()
+    graph.replay()
+
+    for eager_output, graph_output in zip(eager, graph_outputs):
+        assert torch.equal(eager_output, graph_output)
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__]))

test/registered/kernels/ops/moe/test_minimax_quant_scatter.py

@@ -243,6 +243,18 @@ def test_standard_masked_runner_matches_compact_end_to_end(monkeypatch, weight_d
         "use_symmetric_memory",
         lambda *args, **kwargs: nullcontext(),
     )
+    monkeypatch.setattr(
+        deep_gemm_runner.deep_gemm_wrapper,
+        "get_contiguous_layout_alignment",
+        lambda expected_m, num_groups: 32,
+    )
+    monkeypatch.setattr(
+        deep_gemm_runner,
+        "get_exec",
+        lambda: SimpleNamespace(
+            deterministic=SimpleNamespace(enable_deterministic_inference=False)
+        ),
+    )
 
     # UE8M0 packs four 128-wide scale groups into each int32. Use the smallest
     # legal K for both the gate/up and down GEMMs.
@@ -357,17 +369,23 @@ def test_standard_masked_runner_matches_compact_end_to_end(monkeypatch, weight_d
             ).hidden_states,
         )
 
-    compact_is_masked, compact_all_tokens, compact_m_indices, compact_output = (
-        run_with_layout("compact")
-    )
-    masked_is_masked, masked_all_tokens, masked_m_indices, masked_output = (
-        run_with_layout("masked")
-    )
+    (
+        compact_is_masked,
+        compact_all_tokens,
+        compact_m_indices,
+        compact_output,
+    ) = run_with_layout("compact")
+    (
+        masked_is_masked,
+        masked_all_tokens,
+        masked_m_indices,
+        masked_output,
+    ) = run_with_layout("masked")
     torch.cuda.synchronize()
 
     assert not compact_is_masked
     assert masked_is_masked
-    assert compact_all_tokens == 256
+    assert compact_all_tokens == 64
     assert masked_all_tokens is None
     assert masked_m_indices is None
     valid_assignments = topk_ids[topk_ids >= 0]

test/registered/kernels/ops/moe/test_pad_expert_counts.py

@@ -0,0 +1,37 @@
+import unittest
+
+import torch
+
+from sglang.kernels.ops.moe.triton_pad_expert_counts import pad_expert_counts
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cuda_ci(est_time=5, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+
+
+@unittest.skipUnless(torch.cuda.is_available(), "pad_expert_counts needs CUDA")
+class TestPadExpertCounts(CustomTestCase):
+    def test_matches_eager(self):
+        cases = (
+            ([0], 16, 32),
+            ([1, 8, 9, 0], 8, 48),
+            ([0, 1, 7, 8, 9, 31, 32], 16, 160),
+        )
+        for dtype in (torch.int32, torch.int64):
+            for values, block_e, all_tokens in cases:
+                with self.subTest(dtype=dtype, values=values):
+                    counts = torch.tensor(values, device="cuda", dtype=dtype)
+                    expected = (((counts + block_e - 1) // block_e) * block_e).to(
+                        torch.int32
+                    )
+                    expected[-1].add_(all_tokens - expected.sum())
+
+                    actual = pad_expert_counts(counts, block_e, all_tokens)
+
+                    self.assertTrue(torch.equal(actual, expected))
+                    self.assertEqual(actual.dtype, torch.int32)
+                    self.assertEqual(actual.sum().item(), all_tokens)
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/quant/test_fp8_utils.py

@@ -4,17 +4,76 @@ from unittest.mock import patch
 
 import torch
 
+from sglang.srt.layers.quantization.fp8 import (
+    Fp8MoEMethod,
+    _is_cuda,
+    _is_gfx95_supported,
+    _is_hip,
+)
 from sglang.srt.layers.quantization.fp8_utils import (
     inverse_transform_scale_ue8m0,
     quant_weight_ue8m0,
     transform_scale_ue8m0,
 )
+from sglang.srt.runtime_context import get_platform
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=12, stage="base-b", runner_config="1-gpu-large")
 
 
+class TestMxfp8MoeScaleLayout(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        if not (
+            (_is_cuda and get_platform().is_sm100) or (_is_hip and _is_gfx95_supported)
+        ):
+            raise unittest.SkipTest(
+                "MXFP8 MoE quantization requires SM100 or ROCm gfx95"
+            )
+
+    def test_cutlass_serialized_scales_remain_expert_first(self):
+        class CutlassBackend:
+            def is_cutlass(self):
+                return True
+
+            def is_flashinfer_trtllm(self):
+                return False
+
+            def is_flashinfer_trtllm_routed(self):
+                return False
+
+            def is_deep_gemm(self):
+                return False
+
+        layer = SimpleNamespace(
+            w13_weight=torch.nn.Parameter(
+                torch.zeros((2, 64, 32), dtype=torch.float8_e4m3fn, device="cuda")
+            ),
+            w2_weight=torch.nn.Parameter(
+                torch.zeros((2, 32, 32), dtype=torch.float8_e4m3fn, device="cuda")
+            ),
+            w13_weight_scale_inv=torch.nn.Parameter(
+                torch.zeros((2, 64, 1), dtype=torch.uint8, device="cuda"),
+                requires_grad=False,
+            ),
+            w2_weight_scale_inv=torch.nn.Parameter(
+                torch.zeros((2, 32, 1), dtype=torch.uint8, device="cuda"),
+                requires_grad=False,
+            ),
+        )
+        method = object.__new__(Fp8MoEMethod)
+
+        with patch(
+            "sglang.srt.layers.quantization.fp8.get_moe_runner_backend",
+            return_value=CutlassBackend(),
+        ):
+            method._process_mxfp8_moe_weights(layer, quantize=False)
+
+        self.assertEqual(tuple(layer.w13_weight_scale_inv.shape), (2, 64, 1))
+        self.assertEqual(tuple(layer.w2_weight_scale_inv.shape), (2, 32, 1))
+
+
 class TestInverseTransformScaleUe8m0(CustomTestCase):
     def test_round_trip(self):
         for _ in range(100):

test/registered/rl/test_fp32_lm_head.py

@@ -10,6 +10,7 @@ from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.runtime_context import get_context
 from sglang.srt.utils import get_device
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
+from sglang.test.test_utils import CustomTestCase
 
 register_cuda_ci(est_time=9, stage="base-b", runner_config="1-gpu-small")
 register_amd_ci(est_time=15, suite="stage-b-test-1-gpu-small-amd")
@@ -30,7 +31,7 @@ class DummyMeta:
     def compute_dp_attention_metadata(self): ...
 
 
-class TestLMHeadFP32(unittest.TestCase):
+class TestLMHeadFP32(CustomTestCase):
     @classmethod
     def setUpClass(cls):
         if not torch.cuda.is_available() and not (
@@ -38,13 +39,17 @@ class TestLMHeadFP32(unittest.TestCase):
         ):
             raise unittest.SkipTest("needs CUDA GPU or XPU")
 
-    def _make_logprocessor(self, vocab_size, enable_fp32):
+    def _make_logprocessor(self, vocab_size, enable_fp32, config_enable_fp32=False):
         # LogitsProcessor reads get_exec().features.enable_fp32_lm_head
         # from the published config.
         override = get_context().override_server_args(enable_fp32_lm_head=enable_fp32)
         override.install()
         self.addCleanup(override.restore)
-        cfg = SimpleNamespace(vocab_size=vocab_size, final_logit_softcapping=None)
+        cfg = SimpleNamespace(
+            vocab_size=vocab_size,
+            final_logit_softcapping=None,
+            enable_lm_head_fp32=config_enable_fp32,
+        )
         return LogitsProcessor(cfg, skip_all_gather=True, logit_scale=None)
 
     def _run_case(
@@ -55,6 +60,7 @@ class TestLMHeadFP32(unittest.TestCase):
         expected_a_dtype,
         expected_b_dtype,
         expected_operation,
+        config_enable_fp32=False,
     ):
         device = get_device()
         BATCH_SIZE, HIDDEN_SIZE, VOCAB_SIZE = 2, 64, 128
@@ -63,7 +69,9 @@ class TestLMHeadFP32(unittest.TestCase):
         )
         head = LMHeadStub(VOCAB_SIZE, HIDDEN_SIZE, dtype=weights_dtype, device=device)
         meta = DummyMeta()
-        logprocessor = self._make_logprocessor(VOCAB_SIZE, enable_fp32)
+        logprocessor = self._make_logprocessor(
+            VOCAB_SIZE, enable_fp32, config_enable_fp32
+        )
 
         original_matmul = torch.matmul
         original_mm = torch.mm
@@ -173,6 +181,21 @@ class TestLMHeadFP32(unittest.TestCase):
             "matmul",
         )
 
+    def test_model_config_enables_fp32_without_server_flag(self):
+        expected_operation = "mm" if torch.cuda.is_available() else "matmul"
+        expected_dtype = (
+            torch.float32 if expected_operation == "matmul" else torch.bfloat16
+        )
+        self._run_case(
+            torch.bfloat16,
+            False,
+            torch.bfloat16,
+            expected_dtype,
+            expected_dtype,
+            expected_operation,
+            config_enable_fp32=True,
+        )
+
 
 if __name__ == "__main__":
     unittest.main(verbosity=2)

test/registered/unit/configs/test_model_config_scaling.py

@@ -1,5 +1,6 @@
 import math
 import unittest
+from types import SimpleNamespace
 
 from sglang.srt.configs.model_config import ModelConfig, compute_mla_mscale_scaling
 from sglang.test.ci.ci_register import register_cpu_ci
@@ -92,6 +93,31 @@ class TestInitMlaScaling(CustomTestCase):
         base, scaling = _mla_scaling(None)
         self.assertEqual(scaling, base)
 
+    def test_hyv4_shape_derivation_uses_rope_parameters(self):
+        hf_config = SimpleNamespace(
+            architectures=["HYV4ForCausalLM"],
+            model_type="hy_v4",
+            hidden_size=2816,
+            num_hidden_layers=34,
+            num_attention_heads=32,
+            vocab_size=120832,
+            head_dim=64,
+            v_head_dim=256,
+            kv_lora_rank=512,
+            qk_nope_head_dim=192,
+            qk_rope_head_dim=64,
+            index_topk=2048,
+            index_head_dim=128,
+            rope_parameters={"rope_theta": 10_000_000, "rope_type": "default"},
+        )
+        config = ModelConfig.__new__(ModelConfig)
+        config.hf_config = hf_config
+        config.hf_text_config = hf_config
+
+        config._derive_model_shapes()
+
+        self.assertEqual(config.scaling, 1 / math.sqrt(256))
+
 
 if __name__ == "__main__":
     unittest.main()

test/registered/unit/entrypoints/openai/test_serving_chat.py

@@ -800,6 +800,87 @@ class ServingChatTestCase(unittest.TestCase):
 
         self.assertEqual(req.reasoning_effort, "high")
 
+    def test_hunyuan_default_reasoning_effort_is_normalized_for_template(self):
+        self.template_manager.chat_template_name = None
+        self.template_manager.jinja_template_content_format = "string"
+        self.template_manager.reasoning_config = ReasoningToggleConfig(
+            special_case="hunyuan_effort"
+        )
+        self.chat.reasoning_parser = "hunyuan"
+        self.tm.tokenizer.apply_chat_template.return_value = [1, 2, 3]
+
+        cases = [
+            ("none", None, "no_think"),
+            ("none", "xhigh", "high"),
+        ]
+        for default_effort, request_effort, normalized_effort in cases:
+            with self.subTest(
+                default_effort=default_effort, request_effort=request_effort
+            ):
+                self.chat.default_chat_template_kwargs = {
+                    "reasoning_effort": default_effort
+                }
+                req = ChatCompletionRequest(
+                    model="x",
+                    messages=[{"role": "user", "content": "What is 2+2?"}],
+                    reasoning_effort=request_effort,
+                )
+
+                self.chat._process_messages(req, is_multimodal=False)
+
+                kwargs = self.tm.tokenizer.apply_chat_template.call_args.kwargs
+                self.assertEqual(req.reasoning_effort, normalized_effort)
+                self.assertEqual(kwargs["reasoning_effort"], normalized_effort)
+                self.assertNotIn("reasoning_effort", req.chat_template_kwargs)
+
+    def test_hunyuan_reasoning_effort_precedence_survives_conversion(self):
+        self.template_manager.chat_template_name = None
+        self.template_manager.jinja_template_content_format = "string"
+        self.template_manager.reasoning_config = ReasoningToggleConfig(
+            special_case="hunyuan_effort"
+        )
+        self.chat.reasoning_parser = "hunyuan"
+        self.tm.tokenizer.apply_chat_template.return_value = [1, 2, 3]
+
+        cases = [
+            ("xhigh", "none", "high"),
+            (None, "no_think", "no_think"),
+        ]
+        for request_effort, template_effort, normalized_effort in cases:
+            with self.subTest(
+                request_effort=request_effort, template_effort=template_effort
+            ):
+                req = ChatCompletionRequest(
+                    model="x",
+                    messages=[{"role": "user", "content": "What is 2+2?"}],
+                    reasoning_effort=request_effort,
+                    chat_template_kwargs={"reasoning_effort": template_effort},
+                )
+
+                self.chat._convert_to_internal_request(req)
+
+                kwargs = self.tm.tokenizer.apply_chat_template.call_args.kwargs
+                self.assertEqual(req.reasoning_effort, normalized_effort)
+                self.assertEqual(kwargs["reasoning_effort"], normalized_effort)
+                self.assertNotIn("reasoning_effort", req.chat_template_kwargs)
+
+    def test_non_hunyuan_default_reasoning_effort_is_unchanged(self):
+        self.template_manager.chat_template_name = None
+        self.template_manager.jinja_template_content_format = "string"
+        self.tm.tokenizer.apply_chat_template.return_value = [1, 2, 3]
+        self.chat.default_chat_template_kwargs = {"reasoning_effort": "medium"}
+        req = ChatCompletionRequest(
+            model="x",
+            messages=[{"role": "user", "content": "What is 2+2?"}],
+        )
+
+        self.chat._process_messages(req, is_multimodal=False)
+
+        kwargs = self.tm.tokenizer.apply_chat_template.call_args.kwargs
+        self.assertEqual(req.reasoning_effort, "medium")
+        self.assertEqual(kwargs["reasoning_effort"], "medium")
+        self.assertEqual(req.chat_template_kwargs["reasoning_effort"], "medium")
+
     def test_k2_selected_terminator_reaches_sampling_params(self):
         self.tm._config_overrides["reasoning_parser"] = "k2_horizon"
         self.chat = OpenAIServingChat(self.tm, self.template_manager)

test/registered/unit/function_call/test_hunyuan_detector.py

@@ -75,6 +75,17 @@ def _make_tools():
     ]
 
 
+class _Hy4Tokenizer:
+    def get_vocab(self):
+        return {
+            "<tool_calls:opensource>": 1,
+            "<tool_call:opensource>": 2,
+            "<arg_key:opensource>": 3,
+            "<arg_value:opensource>": 4,
+            "<think:opensource>": 5,
+        }
+
+
 class TestHunyuanDetectorHasToolCall(CustomTestCase):
     def setUp(self):
         self.detector = HunyuanDetector()
@@ -111,6 +122,21 @@ class TestHunyuanDetectorDetectAndParse(CustomTestCase):
         self.assertEqual(len(result.calls), 0)
         self.assertEqual(result.normal_text, text)
 
+    def test_hy4_format_without_tool_separator(self):
+        detector = HunyuanDetector(_Hy4Tokenizer())
+        text = (
+            "<tool_calls:opensource>"
+            "<tool_call:opensource>get_weather"
+            "<arg_key:opensource>city</arg_key:opensource>"
+            "<arg_value:opensource>Beijing</arg_value:opensource>"
+            "</tool_call:opensource></tool_calls:opensource>"
+        )
+
+        result = detector.detect_and_parse(text, self.tools)
+
+        self.assertEqual(result.calls[0].name, "get_weather")
+        self.assertEqual(json.loads(result.calls[0].parameters), {"city": "Beijing"})
+
     def test_zero_arg_inline(self):
         text = (
             "<tool_calls><tool_call>get_current_date<tool_sep></tool_call></tool_calls>"
@@ -293,6 +319,43 @@ class TestHunyuanDetectorArgDeserialization(CustomTestCase):
         args = json.loads(result.calls[0].parameters)
         self.assertIs(args["verbose"], True)
 
+    def test_top_level_composed_schema_args(self):
+        cases = (
+            ("anyOf", {"type": "integer"}, "7", 7),
+            ("oneOf", {"type": "boolean"}, "true", True),
+            ("allOf", {"type": "array", "items": {"type": "integer"}}, "[1,2]", [1, 2]),
+        )
+        for keyword, arg_schema, raw_value, expected in cases:
+            with self.subTest(keyword=keyword):
+                function_name = f"composed_{keyword}"
+                tools = [
+                    Tool(
+                        type="function",
+                        function=Function(
+                            name=function_name,
+                            description="Composed schema",
+                            parameters={
+                                keyword: [
+                                    {
+                                        "type": "object",
+                                        "properties": {"value": arg_schema},
+                                    }
+                                ]
+                            },
+                        ),
+                    )
+                ]
+                text = (
+                    f"<tool_calls><tool_call>{function_name}<tool_sep>"
+                    f"<arg_key>value</arg_key><arg_value>{raw_value}</arg_value>"
+                    "</tool_call></tool_calls>"
+                )
+
+                result = self.detector.detect_and_parse(text, tools)
+                args = json.loads(result.calls[0].parameters)
+
+                self.assertEqual(args, {"value": expected})
+
     def test_string_arg_not_deserialized(self):
         """String-typed args should stay as strings even if they look like JSON."""
         text = (
@@ -359,6 +422,23 @@ class TestHunyuanDetectorStreaming(CustomTestCase):
         self.assertEqual(collected[0]["name"], "get_current_date")
         self.assertEqual(json.loads(collected[0]["parameters"]), {})
 
+    def test_hy4_format_without_tool_separator_char_by_char(self):
+        detector = HunyuanDetector(_Hy4Tokenizer())
+        text = (
+            "<tool_calls:opensource>"
+            "<tool_call:opensource>get_weather"
+            "<arg_key:opensource>city</arg_key:opensource>"
+            "<arg_value:opensource>Tokyo</arg_value:opensource>"
+            "</tool_call:opensource></tool_calls:opensource>"
+        )
+        all_calls = []
+        for char in text:
+            all_calls.extend(detector.parse_streaming_increment(char, self.tools).calls)
+
+        collected = _collect_streamed_tool_calls(all_calls)
+        self.assertEqual(collected[0]["name"], "get_weather")
+        self.assertEqual(json.loads(collected[0]["parameters"]), {"city": "Tokyo"})
+
     def test_chunked_tool_call(self):
         detector = self._new_detector()
         chunks = [

test/registered/unit/model_loader/test_modelopt_loader.py

@@ -552,6 +552,21 @@ class TestParseQuantHfConfig(CustomTestCase):
                 self.assertIn("lm_head", quant_config.ignored_layers)
                 self.assertEqual(quant_config.kv_cache_quant_algo, "FP8")
 
+        nested_result = model_config._parse_modelopt_quant_config(
+            {
+                "quantization": {
+                    "quantization": {
+                        "quant_algo": "MXFP8",
+                        "group_size": 32,
+                        "exclude_modules": ["lm_head"],
+                    }
+                }
+            }
+        )
+        self.assertEqual(nested_result["quant_method"], "mxfp8")
+        self.assertEqual(nested_result["scale_fmt"], "ue8m0")
+        self.assertIn("lm_head", nested_result["modules_to_not_convert"])
+
     def test_modelopt_mxfp8_override(self):
         """Generic ModelOpt selection must not route MXFP8 to scalar FP8."""
         self.assertEqual(

test/registered/unit/models/test_hunyuan_v4.py

@@ -0,0 +1,119 @@
+import sys
+from types import SimpleNamespace
+
+import pytest
+import torch
+from torch import nn
+
+from sglang.srt.models import hunyuan_v4
+from sglang.srt.models.deepseek_common.attention_forward_methods import forward_mla
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=4, suite="base-a-test-cpu")
+
+
+def test_attention_gate_uses_attention_tp(monkeypatch):
+    attn_tp_size = 2
+    parallel = SimpleNamespace(
+        attn_tp_rank=min(1, attn_tp_size - 1), attn_tp_size=attn_tp_size
+    )
+    captured = {}
+
+    def fake_attention_init(module, **kwargs):
+        nn.Module.__init__(module)
+        module.hidden_size = kwargs["hidden_size"]
+        module.num_local_heads = kwargs["num_heads"] // parallel.attn_tp_size
+
+    class FakeColumnParallelLinear(nn.Module):
+        def __init__(self, input_size, output_size, **kwargs):
+            super().__init__()
+            captured.update(kwargs)
+            self.output_size_per_partition = output_size // kwargs["tp_size"]
+
+    monkeypatch.setattr(
+        hunyuan_v4.DeepseekV2AttentionMLA, "__init__", fake_attention_init
+    )
+    monkeypatch.setattr(hunyuan_v4, "ColumnParallelLinear", FakeColumnParallelLinear)
+    monkeypatch.setattr(hunyuan_v4, "get_parallel", lambda: parallel)
+    monkeypatch.setattr(
+        hunyuan_v4.HYV4Attention,
+        "_hpc_gated_mla_supported",
+        staticmethod(lambda *args: False),
+    )
+
+    config = SimpleNamespace(
+        rope_parameters={"rope_theta": 10_000, "rope_type": "default"},
+        hidden_size=6144,
+        num_attention_heads=64,
+        qk_nope_head_dim=8,
+        qk_rope_head_dim=4,
+        v_head_dim=256,
+        q_lora_rank=32,
+        kv_lora_rank=16,
+        max_position_embeddings=1024,
+        gating_type="elementwise",
+    )
+
+    attention = hunyuan_v4.HYV4Attention(config, layer_id=0)
+
+    assert captured["tp_rank"] == parallel.attn_tp_rank
+    assert captured["tp_size"] == parallel.attn_tp_size
+    assert attention.local_gate_width == (64 // attn_tp_size) * 256
+    assert attention.linear_gate.output_size_per_partition == attention.local_gate_width
+
+
+class TupleLinear(nn.Module):
+    def __init__(self, input_size, output_size, dtype):
+        super().__init__()
+        self.weight = nn.Parameter(
+            torch.randn(output_size, input_size, dtype=dtype), requires_grad=False
+        )
+
+    def forward(self, inputs):
+        return nn.functional.linear(inputs, self.weight), None
+
+
+def test_attention_gate_non_bf16_model_fallback_parity():
+    torch.manual_seed(0)
+    attention = hunyuan_v4.HYV4Attention.__new__(hunyuan_v4.HYV4Attention)
+    nn.Module.__init__(attention)
+    attention.linear_gate = TupleLinear(8, 256, torch.float32)
+    attention.local_gate_width = 256
+    attention._gate_backend = "eager"
+    attention._gate_fallback_backend = "eager"
+    hidden_states = torch.randn(3, 8)
+    attn_out = torch.randn(3, 256)
+
+    gate = attention.prepare_attention_output_gate(hidden_states)
+    actual = attention.apply_attention_output_gate(attn_out, gate)
+    expected = attn_out * torch.sigmoid(
+        nn.functional.linear(hidden_states, attention.linear_gate.weight)
+    )
+
+    torch.testing.assert_close(actual, expected)
+
+
+def test_prepared_attention_gate_requires_model_application_hook():
+    with pytest.raises(RuntimeError, match="unsigmoided"):
+        forward_mla._apply_attention_output_gate(
+            SimpleNamespace(), torch.ones(1), torch.ones(1)
+        )
+
+
+def test_hpc_attention_gate_is_bf16_only(monkeypatch):
+    fake_hpc = SimpleNamespace(
+        gemm=SimpleNamespace(gated_mla_gemm=object()), __version__="test"
+    )
+    monkeypatch.setitem(sys.modules, "hpc", fake_hpc)
+    monkeypatch.setattr(hunyuan_v4, "get_device_capability", lambda: (10, 0))
+    supported = hunyuan_v4.HYV4Attention._hpc_gated_mla_supported
+    supported.cache_clear()
+    try:
+        assert supported("elementwise", torch.bfloat16, (256, 6144), 256, 6144)
+        assert not supported("elementwise", torch.float32, (256, 6144), 256, 6144)
+    finally:
+        supported.cache_clear()
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__]))

test/registered/unit/models/test_hunyuan_v4_nextn_weight_loading.py

@@ -0,0 +1,42 @@
+import unittest
+from types import SimpleNamespace
+
+import torch
+
+from sglang.srt.models.hunyuan_v4_nextn import HYV4ForCausalLMNextN
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=4, suite="base-a-test-cpu")
+
+
+class TestHunyuanV4NextNWeightLoading(unittest.TestCase):
+    def test_indexer_checkpoint_layout_is_permuted(self):
+        model = object.__new__(HYV4ForCausalLMNextN)
+        model.config = SimpleNamespace(
+            num_hidden_layers=80,
+            index_n_heads=2,
+            index_head_dim=4,
+            qk_rope_head_dim=2,
+        )
+        captured = []
+        object.__setattr__(
+            model,
+            "do_load_weights",
+            lambda weights, **kwargs: captured.extend(weights),
+        )
+        loaded_weight = torch.arange(8).reshape(8, 1)
+
+        model.load_weights(
+            [("model.mtp_layers.0.self_attn.indexer.wq_b.weight", loaded_weight)]
+        )
+
+        self.assertEqual(
+            captured[0][0], "model.layers.80.self_attn.indexer.wq_b.weight"
+        )
+        torch.testing.assert_close(
+            captured[0][1].flatten(), torch.tensor([2, 3, 0, 1, 6, 7, 4, 5])
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/unit/parser/test_hunyuan_reasoning.py

@@ -0,0 +1,43 @@
+import sys
+
+import pytest
+
+from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
+from sglang.srt.parser.hunyuan_reasoning import normalize_hunyuan_reasoning_effort
+from sglang.srt.parser.template_detection import ReasoningToggleConfig
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=1, suite="base-a-test-cpu")
+
+
+@pytest.mark.parametrize(
+    ("effort", "normalized"),
+    [
+        (None, "high"),
+        ("none", "no_think"),
+        ("minimal", "low"),
+        ("low", "low"),
+        ("medium", "high"),
+        ("high", "high"),
+        ("xhigh", "high"),
+        ("max", "high"),
+    ],
+)
+def test_hunyuan_reasoning_effort_normalization(effort, normalized):
+    request = ChatCompletionRequest(
+        model="x",
+        messages=[{"role": "user", "content": "hi"}],
+        reasoning_effort=effort,
+    )
+
+    normalize_hunyuan_reasoning_effort(
+        request,
+        reasoning_parser="hunyuan",
+        reasoning_config=ReasoningToggleConfig(special_case="hunyuan_effort"),
+    )
+
+    assert request.reasoning_effort == normalized
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__]))

test/registered/unit/parser/test_template_manager.py

@@ -269,6 +269,42 @@ class TestTemplateManagerReasoningDetection(unittest.TestCase):
         _, _, parser = self._detect(template, ["<minimax:tool_call>"])
         self.assertEqual(parser, "minimax")
 
+    HYV4_TEMPLATE = (
+        "{%- set reasoning_mode_token = '<｜reasoning_mode:opensource｜>' %}\n"
+        "{%- if not reasoning_effort is defined %}\n"
+        "    {%- set reasoning_effort = 'high' %}\n"
+        "{%- elif reasoning_effort not in ['high', 'low', 'no_think'] %}\n"
+        "{%- endif %}\n"
+        "<tool_call:opensource>{{ name }}<arg_key:opensource>{{ k }}</arg_key:opensource>"
+    )
+
+    HYV4_VOCAB = [
+        "<tool_calls:opensource>",
+        "<tool_call:opensource>",
+        "<arg_key:opensource>",
+        "<arg_value:opensource>",
+    ]
+
+    def test_hyv4_effort_template_detected_with_special_case(self):
+        # Hy4 drops <tool_sep>; detection must key on the effort-mode template
+        # signature plus the suffixed arg tokens instead.
+        force, config, parser = self._detect(self.HYV4_TEMPLATE, self.HYV4_VOCAB)
+
+        self.assertEqual(config, ReasoningToggleConfig(special_case="hunyuan_effort"))
+        self.assertEqual(parser, "hunyuan")
+        self.assertEqual(
+            detect_tool_call_parser(
+                self.HYV4_TEMPLATE, _DummyTokenizer(self.HYV4_VOCAB), config, force
+            ),
+            "hunyuan",
+        )
+
+    def test_hyv4_template_without_arg_tokens_not_hunyuan(self):
+        _, config, parser = self._detect(self.HYV4_TEMPLATE, ["<tool_call:opensource>"])
+
+        self.assertEqual(config, ReasoningToggleConfig(special_case="hunyuan_effort"))
+        self.assertNotEqual(parser, "hunyuan")
+
 
 class TestTemplateDetectionRuleMatrix(unittest.TestCase):
     """Table-driven tests for REASONING_PARSER_RULES and REASONING_MODE_RULES."""

test/registered/unit/server_args/test_server_args.py

@@ -1037,6 +1037,7 @@ class TestContextParallelServerArgs(CustomTestCase):
             moe_dp_size=1,
             ep_size=1,
             pp_size=1,
+            dcp_size=1,
             enable_aiter_allreduce_fusion=False,
         )
         defaults.update(overrides)
@@ -1169,6 +1170,29 @@ class TestContextParallelServerArgs(CustomTestCase):
             resolution_result(args, "dsa_prefill_cp_mode"), "round-robin-split"
         )
 
+    def test_canonical_interleave_cp_mirrors_to_dsa_runtime_aliases(self):
+        server_args = self._new_cp_args(
+            enable_prefill_cp=True,
+            cp_strategy="interleave",
+            attention_backend="dsa",
+        )
+
+        handle_legacy_cp_runtime_compatibility(server_args)
+        handle_context_parallelism(server_args)
+
+        self.assertTrue(
+            resolution_result(server_args, "enable_dsa_prefill_context_parallel")
+        )
+        self.assertFalse(
+            resolution_result(server_args, "enable_prefill_context_parallel")
+        )
+        self.assertEqual(
+            resolution_result(server_args, "dsa_prefill_cp_mode"), "round-robin-split"
+        )
+        self.assertEqual(
+            resolution_result(server_args, "prefill_cp_mode"), "round-robin-split"
+        )
+
     def test_context_parallel_handler_initializes_cp_strategy(self):
         server_args = self._new_cp_args(
             enable_prefill_cp=True,

test/registered/unit/test_model_overrides.py

@@ -1817,8 +1817,8 @@ class TestGoldenModelOverrides(_IsolatedPublish):
             _dsa_split_backend_resolution,
         )
 
-        def _view(arch="DeepseekV32ForCausalLM", **kw):
-            hf = SimpleNamespace(architectures=[arch])
+        def _view(arch="DeepseekV32ForCausalLM", learnable_sink=False, **kw):
+            hf = SimpleNamespace(architectures=[arch], learnable_sink=learnable_sink)
             defaults = dict(
                 kv_cache_dtype="fp8_e4m3",
                 dsa_prefill_backend=None,
@@ -1866,6 +1866,28 @@ class TestGoldenModelOverrides(_IsolatedPublish):
                     "dsa_decode_backend": "flashmla_kv",
                 },
             )
+            for arch in ("HYV4ForCausalLM", "HYV4ForCausalLMNextN"):
+                with self.subTest(arch=arch, backends="default"):
+                    self.assertEqual(
+                        _dsa_split_backend_resolution(
+                            _view(arch=arch, learnable_sink=True)
+                        ),
+                        {
+                            "dsa_prefill_backend": "flashmla_sparse",
+                            "dsa_decode_backend": "flashmla_sparse",
+                        },
+                    )
+                for field, value in (
+                    ("dsa_prefill_backend", "fa3"),
+                    ("dsa_decode_backend", "trtllm"),
+                ):
+                    with self.subTest(arch=arch, field=field, value=value):
+                        with self.assertRaisesRegex(
+                            ValueError, field.replace("_", "-")
+                        ):
+                            _dsa_split_backend_resolution(
+                                _view(arch=arch, learnable_sink=True, **{field: value})
+                            )
             # non-family arch declares nothing
             self.assertEqual(
                 _dsa_split_backend_resolution(_view(arch="LlamaForCausalLM")), {}
@@ -2842,6 +2864,7 @@ class TestGoldenModelOverrides(_IsolatedPublish):
                 prefill_attention_backend=None,
                 decode_attention_backend=None,
                 enable_prefill_cp=False,
+                dcp_size=1,
             )
             defaults.update(kw)
             return SimpleNamespace(**defaults)
@@ -2857,6 +2880,22 @@ class TestGoldenModelOverrides(_IsolatedPublish):
                             _deepseek_family_overrides(_args(), None),
                             {"attention_backend": "dsa", "page_size": 64},
                         )
+                        for arch in ("HYV4ForCausalLM", "HYV4ForCausalLMNextN"):
+                            hf_config = SimpleNamespace(architectures=[arch])
+                            with self.subTest(arch=arch, prefill_cp=True):
+                                with self.assertRaisesRegex(
+                                    ValueError, "--enable-prefill-cp.*HYV4"
+                                ):
+                                    _deepseek_family_overrides(
+                                        _args(enable_prefill_cp=True), hf_config
+                                    )
+                            with self.subTest(arch=arch, dcp_size=2):
+                                with self.assertRaisesRegex(
+                                    ValueError, "--dcp-size > 1.*HYV4"
+                                ):
+                                    _deepseek_family_overrides(
+                                        _args(dcp_size=2), hf_config
+                                    )
                     # HIP without the preshuffle path: page 1
                     with override_platform(is_hip=True):
                         with patch(