Grouped GEMM + dGLU (SM100)
This is an experimental API and subject to change.
JAX support
Canonical MXFP8 (block-scaled, dense weights): the jax.jit-compatible entry point grouped_gemm_dglu_jax_sm100 takes the canonical operands as keywords. They are b_tensor (experts, n, k), already MMA-packed E8M0 sfa_tensor/sfb_tensor bytes (physical or flat), and an explicit norm_const_tensor, along with A (m, k), saved C (m, 2n) BF16/FP16, prob (m,) fp32/bf16, alpha and beta (experts,) fp32, d_dtype e4m3, and act_func "dswiglu", "dgeglu", or "dsituglu". It returns a TupleDict with the eager wrapper’s keys (dbias_tensor and amax_tensor are None). It runs the Rubin dGLU kernel on SM107 and the SM100 kernel otherwise. The activation scalars are compile-time constants of the traced call, and discrete_col_sfd=True packs column scales by expert. Only dprob is zero-initialized, because the kernel accumulates into it; other outputs are not initialized, so rows at or past padded_offsets[-1] are unspecified, as in the Torch path. grouped_gemm_dglu_wrapper_sm100 dispatches JAX inputs with dense b_tensor and sfa_tensor to it; set sf_vec_size=32, an e4m3 d_dtype, and dprob_tensor=None. dbias, discrete weights, deterministic reduction, dynamic scheduling, and round_dgrad_to_input_dtype raise ValueError.
For act_func="dsituglu", the JAX entry supports MXFP8 inputs on SM100/SM103 only. situ_beta1 sets the gate soft cap (default 4.0), and situ_beta2 sets the up-branch soft cap (default 25.0). Both must be finite and positive and are static arguments of the traced JAX call.
Canonical FP4 and BF16 (dense weights): the same entry also takes packed FP4 A/B with sf_vec_size (keyword). Packed FP4 operands travel as uint8 with two values per byte (JAX has no packed FP4 dtype), so K is twice the array extent; their scale buffers are E4M3 with sf_vec_size=16 for NVFP4, or E8M0. FP4 inputs take a BF16/FP16 d_dtype and no norm_const_tensor, and return D_row, dprob and the per-expert amax (experts, 2, 1); D_col is not produced. With a BF16 b_tensor and no scale tensors it runs the BF16 kernel on canonical A (m, k), C (m, 2n) and fp32 prob (m,), returning D_row and dprob. grouped_gemm_dglu_wrapper_sm100 dispatches JAX inputs with any dense b_tensor to the entry.
BF16 backend: supports JAX arrays in discrete weight mode (dswiglu and dgeglu), including generate_dbias=True and caller-provided zero-initialized dprob. Dense BF16 b_tensor and discrete block-scaled weights are not expressible as JAX arrays and raise clear errors. The wrapper is eager, on the CUDA legacy default stream: block_until_ready inputs, synchronize before reading outputs; keep weight arrays alive until the kernel completes.
For jitted BF16 JAX programs use the same entry point grouped_gemm_dglu_jax_sm100 with b_ptrs and n (built on cutlass.jax.cutlass_call; discrete mode): dprob and (with generate_dbias=True) dbias come back as bridge-managed zero-initialized accumulator outputs — no caller-zeroed buffers, no manual synchronization. The padded_offsets values and b_ptrs entries follow the device-data contract in the BF16 section, and the per-expert weight buffers behind b_ptrs must stay alive and unmoved across every execution of the traced computation.
Overview
Unified Grouped GEMM + dGLU fusion: one public class and wrapper select a plain BF16 or legacy block-scaled grouped GEMM fused with a dGLU backward epilogue (dSwiGLU, dGeGLU, or block-scaled dSiTU-GLU) on NVIDIA Blackwell GPUs (SM100/SM103), with a block-scaled dSwiGLU/dGeGLU path on Rubin (SM107). The operation is implemented with CUTLASS/CuTe DSL.
This is a unified API that supports both weight layout modes:
- Dense mode: All expert weights packed into a single contiguous
(N, K, L)tensor - Discrete mode: Per-expert weight pointers (no weight stacking required)
And both backward activation functions:
- dSwiGLU:
act_func="dswiglu"(default) - dGeGLU:
act_func="dgeglu" - dSiTU-GLU:
act_func="dsituglu"(block-scaled SM100/SM103 only)
Groups are contiguous in the M dimension and described by padded_offsets (cumulative aligned end offsets).
Backend dispatch
Mixed families and unsupported pairs are rejected before allocation or
compilation. Each backend’s argument contract is described below.
dsituglu is not available on the BF16 or Rubin backends.
The block-scaled dGeGLU path supports configurable activation alpha, clamp
bounds, and linear offset on both Blackwell and Rubin in dense and discrete
weight modes. These values are constructor configuration and remain part of
the backward wrapper’s compiled-kernel cache key.
BF16 contract
Pass sfa_tensor=None, sfb_tensor=None (or sfb_ptrs=None), and
norm_const_tensor=None; keep sf_vec_size=16, discrete_col_sfd=False, and
epilogue_op=None. BF16 supports finite positive geglu_alpha and finite
ordered glu_clamp_min < glu_clamp_max; defaults are 1.702 and [-7,7].
Use geglu_alpha=1, glu_clamp_max=10, glu_clamp_min=-10, and
linear_offset=0 for V4.1’s clamped SwiGLU. The slope and bounds are compile
constants and participate in plan caching; linear_offset is a runtime scalar.
Tensors, layouts, and equation
For padded rows M, reduction dimension K, compact gradient width N, and
L experts, BF16 uses:
A:(M, K, 1), stride(K, 1, M*K), BF16;- dense
B:(N, K, L), K-major stride(K, 1, N*K), BF16; - discrete
b_ptrs: contiguous CUDA int64 pointers to expert(N, K)BF16 matrices, withn=N,b_dtype=torch.bfloat16, andb_major="k"or"n"; C:(M, 2N, 1), stride(2N, 1, 2M*N), BF16/FP16/FP32;padded_offsets:(L,)int32 cumulative 256-aligned ends;alpha,beta:(L,)FP32;prob:(M, 1, 1)FP32;- caller-zeroed
dprob:(M, 1, 1), stride(1, 1, 1), FP32; D_row:(M, 2N, 1), stride(2N, 1, 2M*N), BF16/FP16/FP32;- caller-zeroed optional
dbias:(L, 2N, 1), stride(2N, 1, 1), BF16.
Torch callers may instead pass the canonical layouts: contiguous A (M, K),
C (M, 2N), prob and dprob (M,), and dense B (L, N, K), or the N-major
(L, K, N) buffer with b_major="n". With 2-D A, the wrapper returns D_row
(M, 2N). The wrapper advertises supports_bf16_canonical_layouts.
padded_offsets values and b_ptrs entries are a device-data contract: the kernel
reads them on device, and neither check_support() nor execute() copies them to
the host (a blocking read would serialize the launch stream and is illegal under
CUDA-graph capture). Malformed values (a decreasing or unaligned offset, a last
offset outside (0, M], a null or misaligned pointer) are
undefined behaviour, as for any raw-pointer interface. Set
CUDNN_FE_GROUPED_GEMM_VALIDATE_DEVICE_VALUES=1 (read once at import) to turn on
blocking debug checks of those values at execute(); that mode raises
RuntimeError when the launch stream is capturing instead of syncing.
For expert g, the compact GEMM gradient and scaled forward activation are
Split X into alternating 32-wide gate/input blocks. For dSwiGLU, with
s = sigmoid(gate):
For dGeGLU, distinguish raw values, clamped activation values, and inclusive clamp derivative masks. With the default slope and bounds:
With the default linear_offset=1, the kernel computes
These masks correct the earlier BF16 implementation, which multiplied by the raw input values inside the clamp interval and therefore did not compute the derivative of the forward activation. Scalar and packed FP32x2 paths are checked against independent FP64 autograd, including values on the boundaries.
dprob accumulates the row sum of the matching unscaled activation times R;
dbias is the per-expert row reduction of interleaved D_row. The
pointer-array tensor is stream-recorded, while every pointed allocation must
remain alive and unchanged until that stream completes.
Class-API execute() requires workspace=, a caller-owned, contiguous,
128-byte-aligned device buffer of at least op.scratch_workspace_bytes() bytes
(never 0); the API allocates nothing and the wrapper allocates it per call on the
launch stream. See the workspace contract in grouped_gemm.md.
The carved scratch byte range must not overlap any live operand. Direct buffer
spans (including packed uint8 FP4 storage and pointer tables) are checked using
host metadata before launch. Disjoint views of one allocation are allowed.
For discrete weights, the caller must also keep the pointed-to weight and scale
allocations disjoint from scratch; execute never reads device pointer tables
back to the host.
sample_padded_offsets may be a metadata-only cudnn.api_base.TensorDesc.
The wrapper return order is exactly d_row_tensor, d_col_tensor,
dprob_tensor, dbias_tensor, amax_tensor, sfd_row_tensor,
sfd_col_tensor. On BF16, d_col_tensor, amax_tensor, sfd_row_tensor, and
sfd_col_tensor are None; dbias_tensor is None unless
generate_dbias=True.
Optional recomputed activation for training
For torch BF16 C and D_row with act_func="dgeglu", the class accepts
sample_activation at construction and requires the corresponding
activation_tensor on every execution. Both parameters default to None.
An undeclared output is rejected at execution. The wrapper accepts the same
caller-allocated activation_tensor and appends it as an eighth result only
when supplied; calls without it retain the original seven-result order.
The output is BF16 with shape (M, N, 1), row-contiguous stride
(N, 1, M*N) modulo unit dimensions, and a 32-byte-aligned data pointer.
For each expert, using the clamped values of beta[expert] * C, it stores
The primal does not depend on the down-projection gradient or alpha; it
remains defined when either the gradient or probability is zero. It can feed
the down-projection weight-gradient GEMM when the forward activation was
discarded. The epilogue reuses its existing clamp and sigmoid calculation and
writes the output directly; it introduces no separate activation kernel.
It writes all addressed padded rows. Capacity after the final expert offset
is unspecified and must not be consumed.
The output must not overlap any input, gradient, metadata, or pointed expert
weight allocation. The API rejects overlapping byte spans of visible tensor
arguments, including views, using host metadata only. With discrete weights,
the caller must also ensure that the allocations behind b_ptrs do not
overlap the output. Output storage must remain alive until the launch stream
finishes. This option currently rejects JAX, block-scaled operands, dSwiGLU,
and non-BF16 C or D_row; their existing default paths are unchanged.
Block-scaled contract
The block-scaled backend performs:
- Block-scaled grouped GEMM: Low-precision GEMM (FP4, FP8) with per-block scale factors across multiple expert groups
- dGLU backward epilogue: Fused backward computation using the forward
Ctensor (input/gate interleaved) - Optional quantized output: Produces row and column scale factors for downstream quantization
Shapes
Canonical layouts
The block-scaled wrapper and API class also accept contiguous A (M, K), C (M, 2N),
prob and dprob (M,), dense B (L, N, K) (or N-major (L, K, N) with b_major="n",
FP8 only), and scale buffers of any rank whose bytes are already MMA-packed. With 2-D A,
the wrapper returns D_row/D_col (M, 2N) and contiguous physical SFD buffers. Legacy
layouts are unchanged. The wrapper advertises supports_canonical_layouts.
Equations
For dSiTU-GLU, define
The fused backward computes
and returns ref * prob * T_u * dT_g/dG and
ref * prob * T_g * dT_u/dU. dprob accumulates the reduction of
ref * T_g * T_u across the output columns in 32-column chunks, producing
shape (valid_m, 1, 1).
The beta values are compile-time specialization values and therefore belong to
the dGLU compiled-kernel cache key.
- Inputs
A: contiguous activation tensor across all groups, shape(valid_m, K, 1)B(dense): weight tensor across all groups, shape(N, K, L)B(discrete): per-expert weight pointers,b_ptrsshape(num_experts,)of int64C: forward intermediate tensor with interleaved input/gate blocks, shape(valid_m, 2N, 1)SFA: scale factor tensor for A, shape(32, 4, ceil(valid_m/128), 4, ceil(ceil(K/sf_vec_size)/4), 1)SFB(dense): scale factor tensor for B, shape(32, 4, ceil(N/128), 4, ceil(ceil(K/sf_vec_size)/4), L)SFB(discrete): per-expert SFB pointers,sfb_ptrsshape(num_experts,)of int64padded_offsets: cumulative sum of aligned group M sizes, shape(L,).valid_m = padded_offsets[-1]alpha: per-group scaling factors for GEMM, shape(L,)beta: per-group scaling factors forCin dSwiGLU, shape(L,); dGeGLU consumes the savedCdirectly and ignoresbetaprob: per-row gating probabilities, shape(valid_m, 1, 1)norm_const: normalization constant for FP8 quantization, shape(1,)
- Outputs
D_row: row-quantized dGLU output, shape(valid_m, 2N, 1)D_col: column-quantized dGLU output, shape(valid_m, 2N, 1)dprob: gradient ofprob, shape(valid_m, 1, 1). Must be zero-initialized.dbias(optional): per-expert bias gradient tensor, shape(L, 2N, 1)SFD_row: row scale factors (whend_dtypeis FP8), shape(32, 4, ceil(valid_m/128), 4, ceil(ceil((2N)/sf_vec_size)/4), 1)SFD_col: column scale factors (whend_dtypeis FP8), shape(32, 4, ceil((2N)/128), 4, ceil(ceil(valid_m/sf_vec_size)/4), 1)amax: per-group amax (whend_dtypeis bf16/float16), shape(L, 2, 1)
Step 1: Block-scaled grouped GEMM (per group g with rows m in [padded_offsets[g-1], padded_offsets[g])):
Step 2: dGLU backward epilogue (performed with 32-column interleaving along 2N):
- Deinterleave
Cinto alternating 32-column gate and up blocks. dSwiGLU appliesbeta_gtoC; dGeGLU uses the saved forward values directly.
For dSwiGLU (act_func="dswiglu"):
swish = gate * sigmoid(gate)dprob += sum(swish * input * ref)over 32-column chunksab = ref * prob * swishdswiglu = ref * prob * input * sigmoid(gate) * (1 + gate * (1 - sigmoid(gate)))- Interleave
[ab, dswiglu]back intoD_row/D_colwith 32-column blocks.
For dGeGLU (act_func="dgeglu"), let gate and up be the original
saved C values and ref the GEMM result from Step 1:
The gate has only an upper clamp; the up branch has both bounds. Clamp masks
use the original values before clamping, and gradients are retained at equality
with either bound. The offset is added after clamping the up branch. The
probability gradient does not include a factor of prob, so it can be nonzero
when the routing probability is zero. [dgate, dup] is stored in alternating
32-column blocks.
The upstream ref includes alpha_tensor[g] ** 2 on both architectures; this
per-group GEMM scale is separate from geglu_alpha. dGeGLU reads C in its
stored precision, so a BF16 saved intermediate is the input to this derivative,
not the original FP32 forward accumulator. C is not modified.
The defaults are geglu_alpha=1.702, glu_clamp_max=7.0,
glu_clamp_min=-7.0, and linear_offset=1.0. For DeepSeek V4 clamped SwiGLU,
use act_func="dgeglu", geglu_alpha=1.0, linear_offset=0.0,
glu_clamp_max=L, and glu_clamp_min=-L, matching the forward configuration
(L=10 for the Flash recipe).
Set these parameters when constructing GroupedGemmDgluSm100 or calling
grouped_gemm_dglu_wrapper_sm100. The backward API specializes the activation
configuration; changing it selects another compiled-kernel cache entry, while
repeating a configuration reuses that entry. It does not expose the forward
API’s per-execution activation controls. With BF16 C, the packed derivative
represents clamp bounds in BF16; limits such as 7 and 10 are exact, while
nonrepresentable limits can differ from the scalar FP32 clamp path.
Step 3: Optional output quantization (when SFD outputs are generated):
Diagram
API usage
BF16
The class API and grouped_gemm_dglu_wrapper_sm100 accept the optional
round_dgrad_to_input_dtype=False compile-time flag. With True, the FP32
GEMM accumulator, after the existing alpha[group]**2 scaling, is rounded to
A’s BF16 dtype and converted back to FP32 before activation derivatives and
dprob reduction. This preserves a materialized BF16 down-projection input
gradient boundary in training. Both scalar and packed dSwiGLU/dGeGLU paths and
dense/discrete weights support it. The default retains FP32 accumulator
precision; C, output casts and accumulator reset requirements are unchanged.
The block-scaled paths reject True. The jitted JAX entry point currently
retains the default and does not expose this flag.
High-level wrapper
Class API
use_dynamic_sched=False selects static scheduling; True caches a dynamic-M
callable for compatible shapes. Cache keys retain compile-sensitive layouts,
dtypes, activation, dbias policy, scheduler, tiles/clusters, features, and
overlap margin.
Block-scaled
High-level wrapper
Dense mode:
Discrete mode:
Class API
Dense mode:
In the class API, dbias generation is specialized at compile time: if sample_dbias is omitted, dbias_tensor must also be omitted at execute().
Discrete mode:
Parameters
Weight Mode
The weight mode is auto-detected from constructor arguments:
- Dense: Provide
sample_bandsample_sfb(contiguous weight tensors) - Discrete: Provide
num_experts,b_shape, andb_dtype(per-expert pointer mode)
Providing both or neither raises ValueError.
Input/Output Tensors
-
Input tensor A:
a_tensor/sample_a- Shape:
(valid_m, K, 1) - Stride:
(K, 1, valid_m·K)— must be K-major - Dtype (
ab_dtype):{float4_e2m1fn_x2, uint8, float8_e4m3fn, float8_e5m2}
- Shape:
-
Input tensor B (dense mode):
b_tensor/sample_b- Shape:
(N, K, L)whereL = num_groups - Stride: K-major or N-major. Must be K-major for FP4.
- Dtype: Must match A
- Shape:
-
Input B pointers (discrete mode):
b_ptrs- Shape:
(num_experts,)— 1-D int64 device tensor of per-expert B data pointers
- Shape:
-
Input tensor C:
c_tensor/sample_c- Shape:
(valid_m, 2N, 1)— forward activation with interleaved input/gate - Stride:
(2N, 1, valid_m·2N)— must be N-major - Dtype:
{float32, float16, bfloat16, float8_e4m3fn, float8_e5m2}
- Shape:
-
Output tensor D_row:
d_row_tensor/sample_d_row- Shape:
(valid_m, 2N, 1) - Stride:
(2N, 1, valid_m·2N)— must be N-major - Dtype (
d_dtype):{bfloat16, float32}for FP4;{float8_e4m3fn, float8_e5m2}for FP8
- Shape:
-
Output tensor D_col:
d_col_tensor/sample_d_col- Shape:
(valid_m, 2N, 1)— must match D_row dtype and stride
- Shape:
-
Input tensor prob:
prob_tensor/sample_prob- Shape:
(valid_m, 1, 1), dtype:float32
- Shape:
-
Output tensor dprob:
dprob_tensor/sample_dprob- Shape:
(valid_m, 1, 1), dtype:float32 - Must be zero-initialized
- Shape:
-
Scaling tensors:
alphashape(L,),betashape(L,),amaxshape(L, 2, 1),norm_constshape(1,)
Common Parameters
acc_dtype: Must betorch.float32mma_tiler_mn: Kernel tile size. Default:(256, 256)TILE_M ∈ {128, 256}TILE_N = 256
cluster_shape_mn: Thread Block cluster shape. Default:(2, 1)whenTILE_M=256,(1, 1)otherwisesf_vec_size: Scale factor vector size.{16, 32}. Default:16vector_f32: Enable packed f32 operations for dSwiGLU and dGeGLU. Default:False. K3-default dSiTU-GLU (situ_beta1=4.0) always uses its packed FP32x2 specialization; non-default dSiTU-GLU uses scalar FP32.m_aligned: Must be256. Default:256discrete_col_sfd: Generate discrete col-major scale factors. Default:Falseact_func: Backward activation function."dswiglu"(default),"dgeglu", or block-scaled"dsituglu"linear_offset: Offset added to the clamped dGeGLU up branch. Default:1.0for dGeGLUgeglu_alpha: dGeGLU sigmoid input scale. Default:1.702glu_clamp_max: dGeGLU upper bound for gate and up. Default:7.0glu_clamp_min: dGeGLU lower bound for up only. Default:-7.0- These activation parameters must match forward; on block-scaled Blackwell and Rubin, they specialize the backward cache. BF16 accepts finite positive
geglu_alphaand finite orderedglu_clamp_min < glu_clamp_max. situ_beta1: Positive finite gate tanh scale for dSiTU-GLU. Default:4.0situ_beta2: Positive finite up-branch tanh scale for dSiTU-GLU. Default:25.0b_major(discrete only): B tensor major dimension."k"(default) or"n". Must be"k"for FP4.epilogue_op: Optional post-processing.None(default),"identity","relu", or"srelu"
Wrapper-specific Parameters
d_dtype: Output D tensor data type. Default:torch.bfloat16cd_major: Must be"n". Default:"n"n(discrete only): B weight N dimensionb_dtype(discrete only): B weight data typedeterministic: Makesdprobbit-exact run to run — see Deterministic dprob. Default:False
Deterministic dprob
By default the N-tiles combine their dprob[token] partials with an fp32 atomic, so dprob can
differ in the last bits between identical runs. With deterministic=True each (token, N-tile)
partial goes to its own fp32 slot, and the slots are summed in a fixed order into dprob on the
launch stream; use_dynamic_sched stays supported and the other outputs are unchanged.
Cost per call: a zeroed valid_m x ceil(N / 256) fp32 workspace (default tile and cluster) and a
two-op reduction. Supported for the SM100 block-scaled kernel (dense or discrete weights) with
torch tensors and without dbias; other configurations raise when the flag is set.
Wrapper Return Values
Returns a TupleDict (dictionary + tuple unpacking):
d_row_tensor: Row-quantized dGLU outputd_col_tensor: Column-quantized dGLU outputdprob_tensor: Gradient of probamax_tensor: Per-group amax (whend_dtypeis bf16/fp16)sfd_row_tensor: Row scale factors (when SFD enabled)sfd_col_tensor: Column scale factors (when SFD enabled)
Support Surface and Constraints
Layouts and Strides
Amust be K-majorBmust be K-major (dense; canonical(L, K, N)withb_major="n"is also accepted) or K/N-major (discrete). Must be K-major for FP4.C,D_row,D_colmust be N-major- All tensors must be 16-byte aligned along the contiguous dimension
Data Types
Additional Type Constraints
AandBmust have the same dtype- Scale factor tensors must have the same dtype
D_rowandD_colmust have the same dtypedbiasmust bebfloat16sf_dtype=float8_e4m3fnis incompatible withsf_vec_size=32- FP8
c_dtypewithvector_f32=Trueis not supported - FP4
ab_dtypeonly supportsd_dtypein{bfloat16, float32} - For non-dbias paths, FP4
ab_dtypewithsf_vec_size=16andd_dtype=float32is not supported - FP8
ab_dtypeonly supportsd_dtypein{float8_e4m3fn, float8_e5m2}
Shapes and Divisibility
Nmust be divisible by 32 (32-column blocks for input/gate interleaving)- Expert count must be
<= 1024 - Each group’s M dimension is aligned to
m_aligned(256) - In the class API,
dbiasis compiled in only whensample_dbiasis provided; passing a runtimedbias_tensorwithoutsample_dbiasraisesValueError use_single_group_runtime_offsets=Trueis supported only by the block-scaled kernel with exactly one expert. In this mode the kernel derivespadded_offsets[0]from runtimeA.shape[0]and does not load its value from device memory; the argument must still be an int32 tensor with shape(1,).
Environment
- Requires CUDA with SM100/SM103 (Blackwell), or SM107 (Rubin) for the block-scaled backend
- Rubin MXFP8 uses matching FP8 A/B operands, E8M0 scale factors, and
sf_vec_size=32; the dGLU API requires FP8 D outputs with row/column scale factors for these inputs
Usage Examples
For usage examples, see test cases in test/python/gemm/cutedsl/test_grouped_gemm_dglu.py (dense mode, unified API) and test/python/gemm/cutedsl/test_discrete_grouped_gemm_dswiglu.py (discrete mode).
Rubin MXFP8 activation-parameter coverage is in
test/python/gemm/cutedsl/test_grouped_gemm_dglu.py
(test_rubin_mxfp8_clamped_dgeglu_*).
The eager JAX wrapper allocates its internal scratch with the CUDA stream-ordered allocator and enqueues its release after the consumer on the same stream. This covers overlapping wrapper calls without storing mutable scratch on a cached plan. Input, pointer-table, weight, and output ownership still follows the eager JAX synchronization contract above. Direct API callers must pass a CUDA workspace on the same device as the operands and keep it alive through completion.