File size: 11,040 Bytes
c83c617 | 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 | """V2 incremental attention with explicit, immutable FP32 cache snapshots.
Caches belong to one request. Reuse only exact token prefixes, with absolute
learned positions; never slide, persist across requests, or mutate a branch.
"""
import math
from pathlib import Path
import numpy as np
import torch
from torch import nn
from vimeml.deployment import v2
from vimeml.deployment.bundle import fresh_directory, read_json, write_json
from vimeml.deployment.coreml import coremltools, inspect_spec
from vimeml.training.data import file_sha
FORMAT = "vimeml_coreml_v2_kv_v1"
class CacheGraph(nn.Module):
def __init__(self, model):
super().__init__()
self.model = model.eval()
c = model.config
self.width, self.heads, self.head_dim = c.d_model, c.n_heads, c.d_model // c.n_heads
mask = torch.full((c.context_length, c.context_length), -float("inf"))
self.register_buffer("causal_mask", torch.triu(mask, diagonal=1))
def forward(self, input_ids, cache_length, key_cache, value_cache):
length = input_ids.shape[1]
past = cache_length[0]
positions = torch.arange(length, device=input_ids.device, dtype=torch.int32) + past
hidden = self.model.token_embedding(input_ids) + self.model.position_embedding(positions)
# Rectangular mask: queries have absolute positions past..end-1.
mask = self.causal_mask[positions.long()]
keys, values = [], []
for i, block in enumerate(self.model.blocks):
normalized = block.attention_norm(hidden)
qkv = block.attention.qkv(normalized).reshape(1, length, 3, self.heads, self.head_dim)
query, key, value = qkv.permute(2, 0, 3, 1, 4).unbind(0)
indices = (positions.long(),)
key = key_cache[i].permute(2, 0, 1, 3).index_put(indices, key.permute(2, 0, 1, 3)).permute(1, 2, 0, 3)
value = value_cache[i].permute(2, 0, 1, 3).index_put(indices, value.permute(2, 0, 1, 3)).permute(1, 2, 0, 3)
attention = torch.softmax(torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(self.head_dim) + mask, dim=-1)
attended = torch.matmul(attention, value).transpose(1, 2).reshape(1, length, self.width)
hidden = hidden + block.attention.projection(attended)
hidden = hidden + block.ffn(block.ffn_norm(hidden))
keys.append(key)
values.append(value)
# Output exactly the new chunk. A one-token decode emits one row;
# avoid a second dynamic output dimension from runtime logit slicing.
logits = self.model.lm_head(self.model.final_norm(hidden))
return logits, torch.stack(keys), torch.stack(values)
def cache_shape(config):
return (config.n_layers, 1, config.n_heads, config.context_length, config.d_model // config.n_heads)
def empty_cache(config):
return tuple(np.zeros(cache_shape(config), dtype=np.float32) for _ in range(2))
@torch.inference_mode()
def trace(model):
graph = CacheGraph(model).eval()
keys, values = (torch.from_numpy(a) for a in empty_cache(model.config))
traced = torch.jit.trace(graph, (torch.tensor([[2, 4, 7]], dtype=torch.int32),
torch.tensor([0], dtype=torch.int32), keys, values))
# Exercise a nonzero past, single-token decode and the full capacity,
# rather than checking only the trace example.
generator = torch.Generator().manual_seed(716)
ids = torch.randint(model.config.vocab_size, (1, model.config.context_length), generator=generator, dtype=torch.int32)
expected = model(ids)
for fn in (graph, traced):
state = (keys, values)
start = 0
for size in (1, 3, model.config.context_length - 4):
actual, k, v = fn(ids[:, start:start+size], torch.tensor([start], dtype=torch.int32),
*state)
torch.testing.assert_close(actual, expected[:, start:start+size], atol=3e-5, rtol=3e-4)
state = k, v
start += size
actual, _, _ = fn(ids, torch.tensor([0], dtype=torch.int32), keys, values)
torch.testing.assert_close(actual, expected, atol=3e-5, rtol=3e-4)
return traced
def finish(output, model, metadata):
fresh_directory(output)
model.save(str(Path(output) / "model.mlpackage"))
files = v2.inventory(Path(output) / "model.mlpackage")
write_json(Path(output) / "manifest.json", {"format": FORMAT, "status": "complete",
"architecture": v2.ARCHITECTURE, **metadata, "environment": v2.environment(),
"spec": inspect_spec(model), "package_files": files,
"package_bytes": sum(item["bytes"] for item in files.values())})
def verify_package(directory):
directory = Path(directory)
m = read_json(directory / "manifest.json")
if (m.get("format") != FORMAT or m.get("status") != "complete"
or m.get("architecture") != v2.ARCHITECTURE or v2.inventory(directory / "model.mlpackage") != m["package_files"]):
raise ValueError("Invalid KV package or changed file inventory.")
return m
def convert(bundle, output):
if Path(output).exists():
raise ValueError("Output exists; choose a new experiment.")
lm, ct = v2.BundleLM(bundle), coremltools()
c = lm.model.config
result = ct.convert(trace(lm.model), source="pytorch", convert_to="mlprogram",
minimum_deployment_target=ct.target.iOS18, compute_precision=ct.precision.FLOAT32,
inputs=[ct.TensorType(name="input_ids", shape=(1, ct.RangeDim(1, c.context_length, default=1)), dtype=np.int32),
ct.TensorType(name="cache_length", shape=(1,), dtype=np.int32),
ct.TensorType(name="key_cache", shape=cache_shape(c), dtype=np.float32),
ct.TensorType(name="value_cache", shape=cache_shape(c), dtype=np.float32)],
outputs=[ct.TensorType(name="logits", dtype=np.float32), ct.TensorType(name="new_key_cache", dtype=np.float32),
ct.TensorType(name="new_value_cache", dtype=np.float32)], skip_model_load=True)
finish(output, result, {"kind": "fp32", "minimum_ios": 18, "bundle": lm.metadata,
"cache_shape": cache_shape(c), "cache_dtype": "float32", "compute_units": "CPU_ONLY",
"cache_bytes_per_snapshot": 2 * int(np.prod(cache_shape(c))) * 4,
"interface": "batch1; absolute learned positions; 0<=past, 1<=Q, past+Q<=128; logits [1,Q,V] for new tokens only. Explicit immutable K/V snapshots.",
"conversion_source_sha256": file_sha(Path(__file__)),
"policy": "No training/weight changes. Request-local exact token prefix reuse; no sliding or cross-request retention."})
def compress(source, output, alignment):
if Path(output).exists():
raise ValueError("Output exists; choose a new experiment.")
m, gate = verify_package(source), read_json(alignment)
if (m["kind"] != "fp32" or not gate.get("passed") or gate.get("format") != "vimeml_kv_alignment_v1"
or gate.get("coreml_manifest_sha256") != file_sha(Path(source) / "manifest.json")):
raise ValueError("A passed alignment report for this exact FP32 KV package is required.")
ct = coremltools()
from coremltools.optimize import coreml as opt
model = ct.models.MLModel(str(Path(source) / "model.mlpackage"), skip_model_load=True)
config = opt.OpLinearQuantizerConfig(mode="linear_symmetric", dtype="int8", granularity="per_block", block_size=32, weight_threshold=2048)
optimization, selection = v2.parameter_selection(opt, model, config)
result = opt.linear_quantize_weights(model, config=optimization)
finish(output, result, {key: m[key] for key in ("minimum_ios", "bundle", "cache_shape", "cache_dtype", "compute_units", "cache_bytes_per_snapshot", "interface", "policy")} | {
"kind": "linear8_fp32_compute", "source_manifest_sha256": file_sha(Path(source) / "manifest.json"),
"fp32_alignment_sha256": file_sha(alignment), "compression": {"method": "linear_symmetric", "bits": 8,
"block_size": 32, "selection": selection}, "conversion_source_sha256": file_sha(Path(__file__))})
class Runtime:
def __init__(self, bundle, directory):
lm = v2.BundleLM(bundle)
self.config, self.metadata = lm.model.config, dict(lm.metadata)
self.manifest = verify_package(directory)
if self.manifest["bundle"]["bundle_manifest_sha256"] != self.metadata["bundle_manifest_sha256"]:
raise ValueError("KV model belongs to another bundle.")
if (self.manifest["cache_shape"] != list(cache_shape(self.config))
or self.manifest["cache_dtype"] != "float32" or self.manifest["compute_units"] != "CPU_ONLY"):
raise ValueError("KV package cache/compute contract differs.")
ct = coremltools()
self.model = ct.models.MLModel(str(Path(directory) / "model.mlpackage"), compute_units=ct.ComputeUnit.CPU_ONLY)
self.metadata.update(precision=self.manifest["kind"], compute_units="CPU_ONLY",
coreml_manifest_sha256=file_sha(Path(directory) / "manifest.json"))
self.calls = self.processed_tokens = 0
def step(self, ids, state=None, past=0, last_only=False):
raw = np.asarray(ids)
if (not raw.size or not np.issubdtype(raw.dtype, np.integer)
or raw.ndim not in (1, 2) or (raw.ndim == 2 and raw.shape[0] != 1)
or (raw < 0).any() or (raw >= self.config.vocab_size).any()):
raise ValueError("Invalid incremental token tensor.")
ids = raw.astype(np.int32).reshape(1, -1)
if (not isinstance(past, int) or past < 0 or past + ids.shape[1] > self.config.context_length
or (state is None and past != 0)):
raise ValueError("Invalid incremental input/cache position.")
state = empty_cache(self.config) if state is None else state
if len(state) != 2 or any(not isinstance(a, np.ndarray) or a.shape != cache_shape(self.config)
or a.dtype != np.float32 or not np.isfinite(a).all() for a in state):
raise ValueError("Invalid cache tensor.")
result = self.model.predict({"input_ids": ids, "cache_length": np.array([past], np.int32),
"key_cache": state[0], "value_cache": state[1]})
self.calls += 1
self.processed_tokens += ids.size
# Own prediction output memory, as in the existing stateless adapter.
# A retained branch snapshot must survive subsequent runtime calls.
logits = np.array(result["logits"], dtype=np.float32, copy=True)
caches = tuple(np.array(result[name], dtype=np.float32, copy=True)
for name in ("new_key_cache", "new_value_cache"))
if (logits.shape != (1, ids.size, self.config.vocab_size)
or not np.isfinite(logits).all()
or any(a.shape != cache_shape(self.config) or not np.isfinite(a).all() for a in caches)):
raise ValueError("Invalid Core ML cache/logits output.")
return (logits[:, -1:].copy() if last_only else logits), caches
|