perf: dispatch linear gemv by decode batch size and unify extension style

- replace the per-shape auto tables in the linear backend with an M-banded rule (M in [2,4] on compute capability 8.0+) that measured at the HBM bandwidth floor across every family, and fold the capability check into the capable guard
- drop the unreachable swiglu auto shape-table machinery so both backends share one env-mode ladder via the new dispatch.env_mode helper
- add __all__ across extension modules, name the rotary registration records, and unify typing to the typing-module style
- rewrite test_linear_dispatch.py around behavioral routing assertions and document the M-banded policy in the developer docs
- Benchmark: L20 SM89, Python dispatch overhead 2.9us to 1.5us, auto now covers every projection shape at M in [2,4].
This commit is contained in:
2026-09-03 07:23:13 +08:00
parent 27abb7c5e7
commit 7540acb43e
14 changed files with 225 additions and 536 deletions
+15 -7
View File
@@ -9,6 +9,8 @@ Layout: x is [batch, seq_len, n_heads, head_dim] (bf16).
freqs_cis is [batch, seq_len, dim/2, 2] (f32) — [cos, sin] pairs.
"""
from typing import Any, Dict, List
import torch
from torch import Tensor
@@ -41,7 +43,7 @@ def _torch_apply(x: Tensor, freqs_cis: Tensor) -> Tensor:
return x_out.to(dtype)
def _rotary_records() -> list:
def _rotary_records() -> List[ImplRecord]:
return [
ImplRecord(
family="rotary",
@@ -61,12 +63,15 @@ def _rotary_records() -> list:
]
register_family(
"rotary",
lambda x, freqs_cis: tensor_axes(x),
_rotary_records,
lambda: _rotary_records()[-1],
)
def _axes(x: Tensor, freqs_cis: Tensor) -> Dict[str, Any]:
return tensor_axes(x)
def _fallback_record() -> ImplRecord:
return _rotary_records()[-1]
register_family("rotary", _axes, _rotary_records, _fallback_record)
def apply_rotary_emb(x: Tensor, freqs_cis: Tensor) -> Tensor:
@@ -80,3 +85,6 @@ def apply_rotary_emb(x: Tensor, freqs_cis: Tensor) -> Tensor:
[batch, seq_len, n_heads, head_dim] (bf16)
"""
return resolve("rotary", x, freqs_cis).record.obj(x, freqs_cis)
__all__ = ["apply_rotary_emb"]