File size: 10,475 Bytes
fc3971e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 | """Unstructured Wanda / SparseGPT on SparseMoE expert Linears.
Shapes are unchanged. Only expert `up_gate_proj` and `down_proj` are pruned.
Layer-sequential: prune block i, then re-forward before scoring block i+1.
"""
from __future__ import annotations
import json
import os
import types
from typing import Any
import torch
import torch.nn as nn
from inference.model.dit.dit_module import SparseMoEMLP
from .sparsegpt_full import SparseGPT
from .wanda_unstructured import WrappedGPT, prune_linear_wanda
FORMAT = "davinci_unstructured_v1"
CAPTURE_ENV = "MAGI_UNSTRUCTURED_CAPTURE"
OUT_ROOT_ENV = "MAGI_UNSTRUCTURED_OUT_ROOT"
METHODS_ENV = "MAGI_UNSTRUCTURED_METHODS"
SPARSITIES_ENV = "MAGI_UNSTRUCTURED_SPARSITIES"
STRIDE_ENV = "MAGI_UNSTRUCTURED_STRIDE"
MAX_ENV = "MAGI_UNSTRUCTURED_MAX"
SOURCE_ENV = "MAGI_UNSTRUCTURED_SOURCE_CKPT"
__all__ = [
"FORMAT",
"moe_expert_linears",
"prune_linear_wanda",
"sequential_unstructured",
"global_mlp_sparsity",
"maybe_attach_capture",
"maybe_prune_captured",
]
def moe_expert_linears(mlp: SparseMoEMLP) -> dict[str, nn.Linear]:
out: dict[str, nn.Linear] = {}
for eid, expert in enumerate(mlp.experts):
out[f"expert{eid}.up_gate_proj"] = expert.up_gate_proj
out[f"expert{eid}.down_proj"] = expert.down_proj
return out
def global_mlp_sparsity(model) -> float:
total = 0
zeros = 0
for layer in model.block.layers:
mlp = getattr(layer, "mlp", None)
if not isinstance(mlp, SparseMoEMLP):
continue
for lin in moe_expert_linears(mlp).values():
w = lin.weight.data
total += int(w.numel())
zeros += int((w == 0).sum())
return zeros / max(total, 1)
def _run_layer(layer, sample: dict[str, Any]):
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
out = layer(
sample["h"],
sample["rope"],
sample["permute_mapping"],
sample["inv_permute_mapping"],
sample["varlen_handler"],
sample["local_attn_handler"],
sample["modality_dispatcher"],
sample["cp_split_sizes"],
)
return out.detach() if torch.is_tensor(out) else out
def _clone_samples(samples: list[dict[str, Any]]) -> list[dict[str, Any]]:
cloned = []
for sample in samples:
item = dict(sample)
item["h"] = sample["h"].clone()
cloned.append(item)
return cloned
@torch.no_grad()
def sequential_unstructured(
model,
samples: list[dict[str, Any]],
sparsity: float,
method: str,
verbose: bool = True,
) -> dict:
method = method.strip().lower()
if method not in ("wanda", "sparsegpt"):
raise ValueError(f"method must be wanda or sparsegpt, got {method}")
# Grouped GEMM calls F.linear on packed weights and skips nn.Linear.forward,
# so Wanda/SparseGPT hooks would see nsamples=0 and zero the expert.
os.environ["MAGI_MOE_GROUPED"] = "off"
work = _clone_samples(samples)
layer_meta: dict[str, Any] = {}
captured_any = False
for layer_idx, layer in enumerate(model.block.layers):
mlp = getattr(layer, "mlp", None)
is_moe = isinstance(mlp, SparseMoEMLP)
linears = moe_expert_linears(mlp) if is_moe else {}
wrappers: dict[str, Any] = {}
handles = []
if is_moe:
if method == "wanda":
wrappers = {name: WrappedGPT(lin) for name, lin in linears.items()}
else:
wrappers = {name: SparseGPT(lin) for name, lin in linears.items()}
def _make_hook(wrapper):
def _hook(_, inp, out):
wrapper.add_batch(inp[0].data, out.data)
return _hook
for name, lin in linears.items():
handles.append(lin.register_forward_hook(_make_hook(wrappers[name])))
for sample in work:
_run_layer(layer, sample)
for handle in handles:
handle.remove()
names_meta = {}
layer_ns = [int(wrappers[name].nsamples) for name in linears]
if sum(layer_ns) == 0:
raise RuntimeError(
f"unstructured {method} layer {getattr(mlp, 'layer_idx', layer_idx)} "
"captured 0 expert activations; MAGI_MOE_GROUPED must be off"
)
captured_any = True
for name, lin in linears.items():
nsamples = int(wrappers[name].nsamples)
if nsamples == 0:
names_meta[name] = {
"sparsity": 0.0,
"shape": list(lin.weight.shape),
"nsamples": 0,
"skipped": True,
}
if hasattr(wrappers[name], "free"):
wrappers[name].free()
continue
if method == "wanda":
sp = prune_linear_wanda(lin, wrappers[name].scaler_row, sparsity)
else:
sp = wrappers[name].fasterprune(sparsity, verbose=verbose)
wrappers[name].free()
names_meta[name] = {
"sparsity": float(sp),
"shape": list(lin.weight.shape),
"nsamples": nsamples,
}
layer_meta[str(int(getattr(mlp, "layer_idx", layer_idx)))] = names_meta
for sample in work:
sample["h"] = _run_layer(layer, sample)
if verbose and is_moe:
print(
f"unstructured {method} layer {getattr(mlp, 'layer_idx', layer_idx)} "
f"sparsity={sparsity:.3f}",
flush=True,
)
torch.cuda.empty_cache()
if not captured_any:
raise RuntimeError(f"unstructured {method} captured 0 expert activations")
return {
"method": method,
"sparsity": float(sparsity),
"keep_ratio": float(1.0 - sparsity),
"global_mlp_sparsity": global_mlp_sparsity(model),
"n_samples": len(samples),
"layers": layer_meta,
}
def maybe_attach_capture(model) -> dict | None:
flag = os.getenv(CAPTURE_ENV, "").strip()
if flag not in ("1", "true", "True", "yes"):
return None
stride = int(os.getenv(STRIDE_ENV, "8") or "8")
max_samples = int(os.getenv(MAX_ENV, "40") or "40")
block = model.block
orig = block.forward
bag: dict[str, Any] = {"samples": [], "calls": 0, "orig": orig}
def wrapped(
self,
x,
rope,
permute_mapping,
inv_permute_mapping,
varlen_handler,
local_attn_handler,
modality_dispatcher,
cp_split_sizes,
):
bag["calls"] += 1
if bag["calls"] % stride == 0 and len(bag["samples"]) < max_samples:
bag["samples"].append(
{
"h": x.detach().clone(),
"rope": rope,
"permute_mapping": permute_mapping,
"inv_permute_mapping": inv_permute_mapping,
"varlen_handler": varlen_handler,
"local_attn_handler": local_attn_handler,
"modality_dispatcher": modality_dispatcher,
"cp_split_sizes": cp_split_sizes,
}
)
return orig(
x,
rope,
permute_mapping,
inv_permute_mapping,
varlen_handler,
local_attn_handler,
modality_dispatcher,
cp_split_sizes,
)
block.forward = types.MethodType(wrapped, block)
model._unstr_capture = bag
print(
f"unstructured capture stride={stride} max={max_samples}",
flush=True,
)
return bag
def save_unstructured_checkpoint(path: str, model, meta: dict) -> None:
payload = {
"format": FORMAT,
"model": model.state_dict(),
"unstructured_meta": meta,
}
os.makedirs(os.path.dirname(os.path.abspath(path)) or ".", exist_ok=True)
torch.save(payload, path)
def maybe_prune_captured(model) -> list[str]:
out_root = os.getenv(OUT_ROOT_ENV, "").strip()
if not out_root:
return []
bag = getattr(model, "_unstr_capture", None)
if not bag or not bag.get("samples"):
raise RuntimeError(
f"{OUT_ROOT_ENV} is set but no TransformerBlock inputs were captured"
)
block = model.block
orig = bag.get("orig")
if orig is not None:
block.forward = orig
methods = [m.strip() for m in os.getenv(METHODS_ENV, "wanda,sparsegpt").split(",") if m.strip()]
sparsities = [
float(x) for x in os.getenv(SPARSITIES_ENV, "0.25,0.50,0.75").split(",") if x.strip()
]
source = os.getenv(SOURCE_ENV, "").strip()
samples = bag["samples"]
print(
f"unstructured prune captured={len(samples)} calls={bag.get('calls')} "
f"methods={methods} sparsities={sparsities}",
flush=True,
)
orig_state = {k: v.detach().cpu().clone() for k, v in model.state_dict().items()}
written: list[str] = []
os.makedirs(out_root, exist_ok=True)
for method in methods:
for sparsity in sparsities:
keep = 1.0 - float(sparsity)
slug = f"{method}_unstr_keep{int(round(keep * 100)):03d}"
out_path = os.path.join(out_root, f"{slug}.pt")
if os.path.isfile(out_path):
print(f"[skip prune] {out_path} exists", flush=True)
written.append(out_path)
continue
print(f"unstructured apply {slug}", flush=True)
model.load_state_dict(orig_state, strict=True)
meta = sequential_unstructured(
model, samples, sparsity=float(sparsity), method=method, verbose=True,
)
meta["source"] = source
meta["captured_calls"] = int(bag.get("calls") or 0)
save_unstructured_checkpoint(out_path, model, meta)
meta_json = os.path.splitext(out_path)[0] + "_meta.json"
with open(meta_json, "w", encoding="utf-8") as f:
json.dump(meta, f, indent=2)
f.write("\n")
print(f"wrote {out_path} global_mlp_sparsity={meta['global_mlp_sparsity']:.4f}", flush=True)
written.append(out_path)
torch.cuda.empty_cache()
return written
|