MCPcopy Create free account
hub / github.com/bitsandbytes-foundation/bitsandbytes / gemv_4bit

Function gemv_4bit

bitsandbytes/functional.py:1300–1334  ·  view source on GitHub ↗
(
    A: Tensor,
    B: Tensor,
    out: Optional[torch.Tensor] = None,
    transposed_A=False,
    transposed_B=False,
    state=None,
)

Source from the content-addressed store, hash-verified

1298
1299
1300def gemv_4bit(
1301 A: Tensor,
1302 B: Tensor,
1303 out: Optional[torch.Tensor] = None,
1304 transposed_A=False,
1305 transposed_B=False,
1306 state=None,
1307):
1308 if state is None:
1309 raise ValueError("state cannot be None. gemv_4bit() requires the state from quantize_4bit()")
1310
1311 absmax = state.absmax
1312 if state.nested:
1313 absmax = dequantize_blockwise(absmax, state.state2) + state.offset
1314
1315 if out is not None:
1316 torch.ops.bitsandbytes.gemv_4bit.out(
1317 A,
1318 B,
1319 state.shape,
1320 absmax,
1321 state.code,
1322 state.blocksize,
1323 out=out,
1324 )
1325 return out
1326
1327 return torch.ops.bitsandbytes.gemv_4bit.default(
1328 A,
1329 B,
1330 state.shape,
1331 absmax,
1332 state.code,
1333 state.blocksize,
1334 )
1335
1336
1337@deprecated("This function is deprecated and will be removed in a future release.", category=FutureWarning)

Callers

nothing calls this directly

Calls 1

dequantize_blockwiseFunction · 0.70

Tested by

no test coverage detected