docs: close v0.6.13 doc-check gaps + fix(moe) misleading topk_indices ICHECK message

csrc/fused_moe/noAuxTcKernels.cu

@@ -348,7 +348,7 @@ void NoAuxTc(TensorView scores, TensorView bias, int64_t n_group, int64_t topk_g
   TVM_FFI_ICHECK(topk_values.dtype() == data_type)
       << "topk_values must have the same dtype as scores";
   TVM_FFI_ICHECK(encode_dlpack_dtype(topk_indices.dtype()) == int32_code)
-      << "topk_indices must have the same dtype as scores";
+      << "topk_indices must be int32 dtype";
 
   // Validate and extract routing_replay_out
   // NOTE: dim0 >= num_tokens is intentionally NOT checked — with CUDA graphs the buffer

docs/api/fused_moe.rst

@@ -28,6 +28,19 @@ Utility Functions
     interleave_moe_scales_for_sm90_mixed_gemm
     fused_topk_deepseek
 
+Multi-LoRA MoE (BGMV)
+---------------------
+
+Batched Gather-Matrix-Vector kernels for serving multiple LoRA adapters on
+top of a Mixture-of-Experts layer (shrink + expand).
+
+.. autosummary::
+    :toctree: ../generated
+
+    bgmv_moe
+    bgmv_moe_shrink
+    bgmv_moe_expand
+
 CUTLASS Fused MoE
 -----------------
 

docs/api/quantization.rst

@@ -94,6 +94,7 @@ importable.
     :toctree: ../generated
 
     nvfp4_quantize_cute_dsl
+    nvfp4_quantize_per_token_cute_dsl
 
 .. currentmodule:: flashinfer.quantization.kernels.mxfp4_quantize
 

flashinfer/quantization/kernels/nvfp4_quantize.py

@@ -1797,7 +1797,52 @@ def nvfp4_quantize_per_token_cute_dsl(
     sf_layout: int = SF_LAYOUT_128x4,
     enable_pdl: bool | None = None,
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-    """Per-token NVFP4 activation quantization using CuTe-DSL."""
+    r"""Per-token NVFP4 activation quantization using the CuTe-DSL kernel.
+
+    Unlike :func:`nvfp4_quantize_cute_dsl`, which applies a single global
+    scale, this variant computes one quantization scale **per row (token)** of
+    the activation. Each row is scaled independently so that its largest
+    magnitude maps to the NVFP4 dynamic range, and the resulting per-token
+    scale is returned alongside the packed FP4 output and the E4M3 block
+    scale factors.
+
+    - E4M3 block scale factors (FP8), ``sf_vec_size = 16``
+    - E2M1 output format (4-bit, 2 values per byte)
+    - Supports 128x4, 8x4, and linear scale-factor layouts
+
+    The kernel is compiled once per ``(K, dtype, sf_layout, pdl)`` tuple and
+    handles varying ``M`` (number of tokens) at runtime without recompilation.
+
+    Parameters
+    ----------
+    input : torch.Tensor
+        2-D activation tensor of shape ``[M, K]`` with dtype fp16/bf16. ``K``
+        must be divisible by ``NVFP4_SF_VEC_SIZE`` (16).
+    global_scale_inv : torch.Tensor
+        Scalar tensor (``float32``) holding the inverse global scale applied on
+        top of the per-token scale. A Python ``float`` is also accepted and
+        wrapped into a tensor internally.
+    sf_layout : int
+        Scale-factor layout (``0=128x4``, ``1=8x4``, ``2=linear``).
+    enable_pdl : bool, optional
+        Whether to enable Programmatic Dependent Launch. Auto-detected from
+        device capability (SM >= 9.0) when ``None``; pass ``False`` to force it
+        off.
+
+    Returns
+    -------
+    Tuple[torch.Tensor, torch.Tensor, torch.Tensor]
+        ``(fp4_output, scale_output, per_token_scale)`` where:
+
+        - ``fp4_output`` is the packed quantized tensor of shape ``[M, K/2]``
+          with dtype ``uint8`` (two E2M1 values per byte).
+        - ``scale_output`` holds the E4M3 block scale factors (``uint8``)
+          reshaped to ``[padded_rows, padded_sf_cols]``. The padding depends on
+          ``sf_layout``: ``linear`` keeps ``M`` rows, while ``128x4`` / ``8x4``
+          pad rows and columns up to the layout tile.
+        - ``per_token_scale`` is the per-row quantization scale of shape
+          ``[M]`` with dtype ``float32``.
+    """
     from ...utils import device_support_pdl
 
     _valid_sf_layouts = (SF_LAYOUT_128x4, SF_LAYOUT_8x4, SF_LAYOUT_LINEAR)