bc_dsa.py imports triton-only kernels unconditionally, so exl3 1.4.x cannot import on a triton-less machine

[pb2025ai-creator]

Summary

exllamav3/modules/attention_fn/bc_dsa.py imports names from dsa_triton at module
scope, unconditionally. dsa_triton.py itself guards the triton import correctly and
defines those names only inside if has_triton:. So on a machine without triton,
import exllamav3 fails — and because architecture/architectures.py imports every
architecture at module level, the failure is not confined to DeepSeek: no model of any
architecture can be loaded.

This matters on Windows in particular, where official triton has no distribution at all.

Reproduction

Windows 11, Python 3.10, torch==2.10.0+cu128, exllamav3==1.4.4+cu128.torch2.10.0,
no triton installed:

>>> from exllamav3 import Config
>>> Config.from_directory("path/to/any-exl3-model")

Import chain:

exllamav3/architecture/architectures.py   (imports every architecture at module level)
  -> deepseek_v4
    -> dsv4
      -> modules/attention_fn/bc_dsa.py:9
         from .dsa_triton import _dsa_attn_split_kernel, ...

dsa_triton.py does the right thing:

try:
    import triton
    import triton.language as tl
    has_triton = True
except ImportError:
    has_triton = False

if has_triton:
    @triton.jit
    def _dsa_attn_split_kernel(...):
        ...

...but bc_dsa.py imports those names regardless, so the guard cannot take effect.

Versions affected

Confirmed present in 1.4.0, 1.4.2, 1.4.3 and 1.4.4.

Suggested fix

Guard the import in bc_dsa.py the same way dsa_triton.py guards its own, and defer the
failure to the point of use — so a machine without triton can still load architectures that
do not need it. Something like:

try:
    from .dsa_triton import _dsa_attn_split_kernel, ...
    _HAS_DSA_TRITON = True
except ImportError:
    _HAS_DSA_TRITON = False

...with a clear error raised only when a DSA path is actually taken.

Workaround

Installing triton-windows supplies a triton module and the import succeeds; with that
in place 1.4.4 loads and generates normally here (23 EXL3 model directories, text and
vision, including the vision tower path).

⚠ One caveat worth recording for other Windows users: with triton-windows present,
additionally installing the declared dependency flash-linear-attention makes
import exllamav3 fail again, this time inside fla/ops/simple_gla/parallel.py under
triton's JIT source parser (observed at triton-windows 3.5.1 and 3.7.1). Omitting
flash-linear-attention works because every fla import inside exllamav3 is guarded;
a GatedDeltaNet model (Qwen3.5-9B) loads and generates without it via gdn.cu.