[Kernel] Triton implementation of causal-conv1d for Mamba-based models (#18218)

Signed-off-by: Tuan M. Hoang-Trong <tmhoangt@us.ibm.com>
Co-authored-by: Tuan M. Hoang-Trong <tmhoangt@us.ibm.com>
Co-authored-by: Tyler Michael Smith <tysmith@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Tuan, Hoang-Trong
2025-07-09 12:53:55 -07:00
committed by GitHub
co-authored by Tuan M. Hoang-Trong Tyler Michael Smith Tyler Michael Smith
parent 31b96d1c64
commit 47043eb678
15 changed files with 1117 additions and 1142 deletions
+5 -29
View File
@@ -963,17 +963,17 @@ def cutlass_fp4_moe_mm(a_tensors: torch.Tensor, b_tensors: torch.Tensor,
expert_offsets: torch.Tensor, sf_offsets: torch.Tensor,
out_dtype: torch.dtype, device: torch.device):
"""
An FP4 Blockscaled Group Gemm that takes in a_tensors, b_tensors and runs
An FP4 Blockscaled Group Gemm that takes in a_tensors, b_tensors and runs
the gemms for each combination based on the specified problem sizes.
This is used as the MoE gemm during NVFP4 Quantized FusedMoE forward.
- a/b_tensors: the NVFP4 a_ptrs and b_ptrs tensors which are quantized
input and expert weights.
- a_/b_scales: The blockscales in FP8-E4M3 precision
- expert_offsets/sf_offsets: Indices that mark at which token index
each expert begins its computation. The number of tokens
computed with expert E is expert_offsets[E + 1] -
expert_offsets[E] And the sf_size per expert is
- expert_offsets/sf_offsets: Indices that mark at which token index
each expert begins its computation. The number of tokens
computed with expert E is expert_offsets[E + 1] -
expert_offsets[E] And the sf_size per expert is
sf_offset[E+1] - sf_offset[E]
- problem_sizes: MxNxK sizes of each expert's multiplication in two grouped
MMs used in the fused MoE operation.
@@ -1464,30 +1464,6 @@ def ggml_moe_get_block_size(quant_type: int) -> int:
# mamba
def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
bias_: Optional[torch.Tensor],
conv_states: Optional[torch.Tensor],
query_start_loc: Optional[torch.Tensor],
cache_indices: Optional[torch.Tensor],
has_initial_state: Optional[torch.Tensor],
silu_activation: bool, pad_slot_id: int):
torch.ops._C.causal_conv1d_fwd(x, weight, bias_, conv_states,
query_start_loc, cache_indices,
has_initial_state, silu_activation,
pad_slot_id)
def causal_conv1d_update(x: torch.Tensor, conv_state: torch.Tensor,
weight: torch.Tensor, bias_: Optional[torch.Tensor],
silu_activation: bool,
cache_seqlens: Optional[torch.Tensor],
conv_state_indices: Optional[torch.Tensor],
pad_slot_id: int):
torch.ops._C.causal_conv1d_update(x, conv_state, weight, bias_,
silu_activation, cache_seqlens,
conv_state_indices, pad_slot_id)
def selective_scan_fwd(u: torch.Tensor, delta: torch.Tensor, A: torch.Tensor,
B: torch.Tensor, C: torch.Tensor,
D_: Optional[torch.Tensor], z_: Optional[torch.Tensor],