[BUG] CuTe DSL from_dlpack rejects float8_e4m3fn tensors (dtype_map lists Float8E5M2 only) #3716

Describe the bug

cutlass.cute.runtime.from_dlpack rejects torch.float8_e4m3fn tensors:

KeyError: Unsupported element type in dlpack: 'Float8E4M3FN'. Supported types are:
['Bool', 'Int64', 'Int32', 'Int16', 'Int8', 'UInt64', 'UInt32', 'UInt16', 'UInt8',
 'Float64', 'Float32', 'Float16', 'BFloat16', 'Float8E5M2']

The runtime already decodes the DLPack type code and reports the element type as
Float8E4M3FN (the string comes out of libcute_dsl_runtime.so, which also knows
Float8E4M3, Float8E4M3FNUZ, Float8E8M0FNU, Float4E2M1FN, ...). Only the Python
lookup table in TensorDescriptor.element_type /
python/CuTeDSL/cutlass/base_dsl/runtime/tensor_descriptor.py stops at Float8E5M2, so
every other 8-bit float that the runtime recognizes is rejected at import time even though
cutlass.Float8E4M3FN is a fully supported kernel element type.

PyTorch is not the limiting side: torch.empty(8, dtype=torch.float8_e4m3fn).__dlpack__()
succeeds on torch 2.11 and the capsule round-trips through torch.from_dlpack.

Steps/Code to reproduce bug

import torch
from cutlass.cute.runtime import from_dlpack

x = torch.empty(8, dtype=torch.float8_e4m3fn, device="cuda")
t = from_dlpack(x, assumed_align=16)   # KeyError: Unsupported element type in dlpack: 'Float8E4M3FN'

Expected behavior

from_dlpack returns a cute.Tensor with element_type == cutlass.Float8E4M3FN, the same
way it does for torch.float8_e5m2.

Workaround in use

Callers currently export the storage as uint8/int8 and overwrite element_type
afterwards, e.g. FlashInfer does this in flashinfer/gemm/kernels/bmm_fp8_blackwell.py,
flashinfer/attention/cute_dsl/fmha.py, and pending PR flashinfer-ai/flashinfer#5533:

t = from_dlpack(x.view(torch.uint8), assumed_align=16)
t.element_type = cutlass.Float8E4M3FN

Adding the missing names to the dtype_map (at least Float8E4M3FN; ideally the other
strings the runtime emits) would let these call sites drop the reinterpretation.

Environment details

• nvidia-cutlass-dsl 4.8.0 (nvidia_cutlass_dsl_libs_cu13-4.8.0), and main as of 2026-10-07
• torch 2.11.0+cu130
• Linux x86_64, CUDA 13.0