Files
2026-07-13 12:34:46 +08:00

468 lines
16 KiB
Python

#!/usr/bin/env python3
"""
SplitLinear: Drop-in replacement for nn.Linear / nn.QuantizedLinear.
Splits GEMM along output channels — ANE does ~65%, GPU does ~35%, concurrent.
Key optimization: same-input projections (Q/K/V, Gate/Up) share a single
input preparation via _InputGroup, eliminating redundant transpose+eval+numpy.
Usage:
from split_linear import patch_model
bridge = patch_model(model, seq=512)
API:
patch_model(model, seq) → high-level one-liner
SplitLinear(layer, bridge, seq) → single layer replacement
ANEBridge() → ANE private API wrapper
"""
import os, ctypes
from concurrent.futures import ThreadPoolExecutor
import numpy as np
import mlx.core as mx
import mlx.nn as nn
LIB_DIR = os.path.dirname(os.path.abspath(__file__))
SPLIT_ALIGN = 64
MIN_SEQ_FOR_SPLIT = 192 # Below this, split overhead > benefit
# ─── ANE Bridge ───
class ANEBridge:
"""Thin wrapper around ANE private API."""
_instance = None
def __init__(self):
lib = ctypes.CDLL(os.path.join(LIB_DIR, 'libane_bridge_v6.dylib'))
lib.ane_init.restype = ctypes.c_int
lib.ane_load_model.restype = ctypes.c_int
lib.ane_load_model.argtypes = [ctypes.c_int, ctypes.c_int, ctypes.c_int,
ctypes.POINTER(ctypes.c_float)]
lib.ane_run.restype = ctypes.c_int
lib.ane_run.argtypes = [ctypes.c_int, ctypes.POINTER(ctypes.c_float),
ctypes.POINTER(ctypes.c_float)]
lib.ane_run_rowmajor.restype = ctypes.c_int
lib.ane_run_rowmajor.argtypes = [ctypes.c_int, ctypes.POINTER(ctypes.c_float),
ctypes.c_int, ctypes.POINTER(ctypes.c_float)]
lib.ane_model_count.restype = ctypes.c_int
lib.ane_surface_count.restype = ctypes.c_int
assert lib.ane_init() == 0, "ANE init failed"
self.lib = lib
def load(self, ic, oc, seq, w_fp32):
w = np.ascontiguousarray(w_fp32, dtype=np.float32)
h = self.lib.ane_load_model(ic, oc, seq,
w.ctypes.data_as(ctypes.POINTER(ctypes.c_float)))
assert h >= 0, f"ANE load failed: {ic}{oc}, seq={seq}"
return h
def run(self, h, inp, out):
self.lib.ane_run(h,
inp.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
out.ctypes.data_as(ctypes.POINTER(ctypes.c_float)))
def run_rowmajor(self, h, inp_rm, L, out_rm):
"""ANE compute with row-major I/O. Uses vDSP for transpose.
inp_rm: [L, IC] row-major float32
out_rm: [L, OC] row-major float32 (pre-allocated)
"""
self.lib.ane_run_rowmajor(h,
inp_rm.ctypes.data_as(ctypes.POINTER(ctypes.c_float)),
L,
out_rm.ctypes.data_as(ctypes.POINTER(ctypes.c_float)))
@property
def model_count(self):
return self.lib.ane_model_count()
@classmethod
def shared(cls):
if cls._instance is None:
cls._instance = cls()
return cls._instance
# Single ANE worker thread
_ane_pool = ThreadPoolExecutor(max_workers=1, thread_name_prefix='ane')
# ─── Helpers ───
def _extract_weight(layer):
"""Extract FP32 weight [OC, IC] from nn.Linear or nn.QuantizedLinear."""
if isinstance(layer, nn.QuantizedLinear):
mx.eval(layer.weight, layer.scales)
b = getattr(layer, 'biases', None)
if b is not None:
mx.eval(b)
w = mx.dequantize(layer.weight, layer.scales, b,
layer.group_size, layer.bits)
else:
w = layer.weight
mx.eval(w)
return np.array(w, dtype=np.float32)
class _InputGroup:
"""
Shared input preparation for same-input projections (Q/K/V or Gate/Up).
Zero-copy path: eval x as FP32 row-major, share numpy view across group.
No GPU transpose needed — ane_run_rowmajor handles transpose in C via vDSP.
"""
__slots__ = ('_np', '_x2d', '_L', '_seq')
def __init__(self, ic=0, seq=0):
self._np = None
self._x2d = None
self._L = 0
self._seq = seq
def prepare(self, x):
"""Prepare input. Called by first proj in group.
Returns (inp_np_rowmajor [L, IC], x_2d, L)."""
x_2d = x.reshape(-1, x.shape[-1]) if x.ndim == 3 else x
L = x_2d.shape[0]
# Cast to FP32 contiguous row-major, eval, then zero-copy numpy view
x_f32 = x_2d.astype(mx.float32) if x_2d.dtype != mx.float32 else x_2d
x_f32 = mx.contiguous(x_f32)
mx.eval(x_f32)
self._np = np.array(x_f32, copy=False) # [L, IC] row-major, zero-copy
self._L = L
self._x2d = x_2d
return self._np, x_2d, L
def get(self):
"""Get cached input. Called by subsequent projs.
Returns (inp_np_rowmajor [L, IC], x_2d, L)."""
return self._np, self._x2d, self._L
# ─── SplitLinear ───
class SplitLinear:
"""
Drop-in nn.Linear replacement with ANE+GPU tensor parallelism.
Two modes controlled externally via set_prefill(True/False):
- prefill: ANE ~65% + GPU ~35% concurrent
- decode: original nn.Linear on GPU → zero overhead
"""
_prefill_mode = False # class-level flag
def __init__(self, layer, bridge, seq, ane_frac=None, name="",
input_group=None, is_first=True):
self.name = name
self._orig = layer # always keep original for decode
w_np = _extract_weight(layer)
oc, ic = w_np.shape
self.ic = ic
self.oc = oc
self.seq = seq
self._is_first = is_first
# Auto-detect split fraction
if ane_frac is None:
if seq < MIN_SEQ_FOR_SPLIT:
ane_frac = 0.0
elif ic > oc * 2:
ane_frac = 0.0 # Wide→narrow: ANE inefficient (down_proj)
elif oc < SPLIT_ALIGN * 2:
ane_frac = 0.0
else:
ane_frac = 0.65
self.ane_frac = ane_frac
if ane_frac <= 0:
self.mode = 'gpu'
return
self.mode = 'split'
self.ane = bridge
self.ane_oc = (int(oc * ane_frac) // SPLIT_ALIGN) * SPLIT_ALIGN
self.gpu_oc = oc - self.ane_oc
if self.ane_oc < SPLIT_ALIGN or self.gpu_oc < 1:
self.mode = 'gpu'
return
self.h_ane = bridge.load(ic, self.ane_oc, seq, w_np[:self.ane_oc, :])
self.buf_ane = np.empty((seq, self.ane_oc), dtype=np.float32) # ANE FP32 output
self.buf_ane_f16 = np.empty((seq, self.ane_oc), dtype=np.float16) # FP16 cast buffer
# GPU weight for split path
self._is_quantized = isinstance(layer, nn.QuantizedLinear)
if self._is_quantized:
# Re-quantize GPU portion as QuantizedLinear → native quantized_matmul
w_gpu_fp32 = mx.array(w_np[self.ane_oc:, :]) # [gpu_oc, ic] float32
w_q, scales, biases = mx.quantize(w_gpu_fp32,
group_size=layer.group_size,
bits=layer.bits)
self._gpu_layer = nn.QuantizedLinear(
ic, self.gpu_oc, bias=False,
group_size=layer.group_size, bits=layer.bits)
self._gpu_layer.weight = w_q
self._gpu_layer.scales = scales
self._gpu_layer.biases = biases
mx.eval(self._gpu_layer.parameters())
self._w_gpu = None # not used for quantized path
else:
self._gpu_layer = None
self._w_gpu = None
if input_group is not None:
self._grp = input_group
else:
self._grp = _InputGroup(ic, seq)
self._is_first = True
@classmethod
def set_prefill(cls, enabled):
cls._prefill_mode = enabled
def __call__(self, x):
# decode / gpu-only / short seq: use original nn.Linear
if self.mode == 'gpu' or not SplitLinear._prefill_mode:
return self._orig(x)
L = x.shape[-2] if x.ndim == 3 else x.shape[0]
if L < MIN_SEQ_FOR_SPLIT:
return self._orig(x)
# ── prefill split path (zero-copy + concurrent ANE/GPU) ──
orig_shape = x.shape
# Input preparation (shared or fresh) — row-major [L, IC] FP32
if self._is_first:
inp_np, x_2d, L = self._grp.prepare(x)
else:
inp_np, x_2d, L = self._grp.get()
# ANE: launch in worker thread (row-major path, vDSP transpose in C)
out_buf = self.buf_ane[:L] # [L, ane_oc] view
fut = _ane_pool.submit(self.ane.run_rowmajor, self.h_ane, inp_np, L, out_buf)
# GPU: matmul with GPU portion weight
if self._gpu_layer is not None:
# Quantized: use native quantized_matmul via QuantizedLinear
gpu_out = self._gpu_layer(x_2d)
elif self._w_gpu is not None:
gpu_out = x_2d @ self._w_gpu.T
else:
w_gpu = self._orig.weight[self.ane_oc:, :]
gpu_out = x_2d @ w_gpu.T
mx.eval(gpu_out) # sync GPU — enables concurrent ANE execution
# Wait for ANE
fut.result()
# Merge: numpy FP32→FP16 cast + lazy concat
np.copyto(self.buf_ane_f16[:L], out_buf, casting='same_kind')
ane_out = mx.array(self.buf_ane_f16[:L]) # [L, ane_oc] FP16
merged = mx.concatenate([ane_out, gpu_out], axis=-1)
if len(orig_shape) == 3:
merged = merged.reshape(orig_shape[0], orig_shape[1], -1)
return merged
def __repr__(self):
if self.mode == 'gpu':
return f"SplitLinear({self.name}, gpu_only, {self.ic}{self.oc})"
return (f"SplitLinear({self.name}, {self.ane_oc}ane+{self.gpu_oc}gpu, "
f"{self.ic}{self.oc}, frac={self.ane_frac:.0%})")
# ─── Tree Walk ───
def _find_linears(module, prefix=""):
"""Walk MLX model tree, yield (parent, key, full_name, linear)."""
if not hasattr(module, 'children'):
return
for attr_name, child in module.children().items():
full_name = f"{prefix}.{attr_name}" if prefix else attr_name
if isinstance(child, (nn.Linear, nn.QuantizedLinear)):
yield (module, attr_name, full_name, child)
elif isinstance(child, nn.Module):
yield from _find_linears(child, full_name)
elif isinstance(child, list):
for i, v in enumerate(child):
fname = f"{full_name}.{i}"
if isinstance(v, (nn.Linear, nn.QuantizedLinear)):
yield (child, i, fname, v)
elif hasattr(v, 'children'):
yield from _find_linears(v, fname)
elif isinstance(child, dict):
for k, v in child.items():
fname = f"{full_name}.{k}"
if isinstance(v, (nn.Linear, nn.QuantizedLinear)):
yield (module, k, fname, v)
elif hasattr(v, 'children'):
yield from _find_linears(v, fname)
elif hasattr(child, 'children'):
yield from _find_linears(child, full_name)
def _get_input_dims(layer):
"""Get true input dimensions for nn.Linear or nn.QuantizedLinear."""
if isinstance(layer, nn.QuantizedLinear):
return layer.weight.shape[-1] * 32 // layer.bits
return layer.weight.shape[-1]
def patch_model(model, seq, bridge=None, verbose=True):
"""
Patch all linear layers with SplitLinear.
Same-input projections share an InputGroup:
- Q/K/V share one (same hidden_state input)
- Gate/Up share one (same hidden_state after attn)
- O and Down each get their own (unique inputs)
Returns: bridge instance.
"""
if bridge is None:
bridge = ANEBridge.shared()
lang = model.language_model
lm = lang.model
N = len(lm.layers)
n_split = n_gpu = 0
for li in range(N):
la = lm.layers[li]
attn = la.self_attn
mlp = la.mlp
ic_attn = _get_input_dims(attn.q_proj)
ic_mlp = _get_input_dims(mlp.gate_proj)
# Q/K/V share input group
qkv_grp = _InputGroup(ic_attn, seq)
for i, name in enumerate(('q_proj', 'k_proj', 'v_proj')):
orig = getattr(attn, name)
sl = SplitLinear(orig, bridge, seq,
name=f"layer.{li}.{name}",
input_group=qkv_grp,
is_first=(i == 0))
setattr(attn, name, sl)
n_split += 1 if sl.mode == 'split' else 0
n_gpu += 1 if sl.mode == 'gpu' else 0
# O proj — own input
sl = SplitLinear(attn.o_proj, bridge, seq,
name=f"layer.{li}.o_proj")
attn.o_proj = sl
n_split += 1 if sl.mode == 'split' else 0
n_gpu += 1 if sl.mode == 'gpu' else 0
# Gate/Up share input group
gu_grp = _InputGroup(ic_mlp, seq)
for i, name in enumerate(('gate_proj', 'up_proj')):
orig = getattr(mlp, name)
sl = SplitLinear(orig, bridge, seq,
name=f"layer.{li}.{name}",
input_group=gu_grp,
is_first=(i == 0))
setattr(mlp, name, sl)
n_split += 1 if sl.mode == 'split' else 0
n_gpu += 1 if sl.mode == 'gpu' else 0
# Down proj — own input (auto GPU-only due to IC>OC*2)
sl = SplitLinear(mlp.down_proj, bridge, seq,
name=f"layer.{li}.down_proj")
mlp.down_proj = sl
n_split += 1 if sl.mode == 'split' else 0
n_gpu += 1 if sl.mode == 'gpu' else 0
if verbose:
print(f"[SplitLinear] {N} layers: {n_split} split, {n_gpu} gpu-only")
print(f"[SplitLinear] ANE models: {bridge.model_count}")
return bridge
# ─── Self-test ───
if __name__ == '__main__':
import sys, time
from mlx_vlm.utils import load as vlm_load
from mlx_vlm.models.cache import KVCache
from mlx.utils import tree_flatten
SEQ = int(sys.argv[1]) if len(sys.argv) > 1 else 512
N_W = 3; N_B = 10
FP16_MODEL = '~/Downloads/weights/mlx/Qwen3-VL-2B-Instruct-16bit'
print(f"\n{'='*60}")
print(f" SplitLinear Self-Test (seq={SEQ})")
print(f"{'='*60}\n")
model, _ = vlm_load(FP16_MODEL)
# Cast to true FP16
flat = tree_flatten(model.trainable_parameters())
fp16 = [(k, v.astype(mx.float16)) for k, v in flat]
model.load_weights(fp16)
mx.eval(model.parameters())
lang = model.language_model
N = lang.args.num_hidden_layers
ids = mx.ones((1, SEQ), dtype=mx.int32)
pos = mx.broadcast_to(mx.arange(SEQ).reshape(1, SEQ)[None, :, :], (3, 1, SEQ))
mx.eval(ids, pos)
# GPU baseline
print("[1/3] GPU FP16 baseline")
for _ in range(N_W):
c = [KVCache() for _ in range(N)]
mx.eval(lang(ids, cache=c, position_ids=pos).logits)
ts = []
for _ in range(N_B):
c = [KVCache() for _ in range(N)]
t0 = time.perf_counter()
mx.eval(lang(ids, cache=c, position_ids=pos).logits)
ts.append((time.perf_counter()-t0)*1000)
bl = float(np.median(ts))
print(f" {bl:.1f}ms\n")
# Reference logits (FP32 for cos_sim)
c_ref = [KVCache() for _ in range(N)]
ref = np.array(lang(ids, cache=c_ref, position_ids=pos).logits.astype(mx.float32))
# Patch
print("[2/3] Patch + benchmark")
bridge = patch_model(model, SEQ)
SplitLinear.set_prefill(True)
for _ in range(N_W):
c = [KVCache() for _ in range(N)]
mx.eval(lang(ids, cache=c, position_ids=pos).logits)
# Accuracy
c_hyb = [KVCache() for _ in range(N)]
hyb = np.array(lang(ids, cache=c_hyb, position_ids=pos).logits.astype(mx.float32))
cos = float(np.dot(ref.flatten(), hyb.flatten()) /
(np.linalg.norm(ref.flatten()) * np.linalg.norm(hyb.flatten()) + 1e-12))
top1 = float((ref.argmax(-1) == hyb.argmax(-1)).mean() * 100)
print(f" cos={cos:.6f}, top1={top1:.1f}%")
# Benchmark
ts = []
for i in range(N_B):
c = [KVCache() for _ in range(N)]
t0 = time.perf_counter()
mx.eval(lang(ids, cache=c, position_ids=pos).logits)
t = (time.perf_counter()-t0)*1000
ts.append(t)
print(f" Run {i+1}: {t:.1f}ms")
med = float(np.median(ts))
print(f"\n{'='*60}")
print(f" GPU FP16: {bl:.1f}ms")
print(f" SplitLinear: {med:.1f}ms ({bl/med:.3f}x)")
print(f" cos={cos:.6f} top1={top1:.1f}%")
print(f" delta: {med - bl:+.1f}ms")
print(f"{'='*60}")