File size: 10,795 Bytes
62274c6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""CPU tests for dtype-aware PLE mmap gathering and checkpoint validation.

Run inside the vLLM image (needs numpy + torch, no GPU):
  docker run --rm -v $PWD:/t -w /t --entrypoint python3 vllm/vllm-openai:qwen38-flash-next test_ple_mmap_cpu.py
"""
import json
import os
import struct
import sys
import tempfile
import time

import numpy as np
import torch

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import vllm_ple_mmap as m  # noqa: E402
import patch_qwen4_exp_config as config_compat  # noqa: E402
import patch_qwen4_exp_quantized_lm_head as lm_head_compat  # noqa: E402
import patch_parallel_lm_head_linear_attrs as lm_head_attrs  # noqa: E402

source_fixture = """\
class QwenTextConfig:
    def __init__(self, layer_types=None, **kwargs):
        super().__init__(layer_types=layer_types, **kwargs)
"""
patched_fixture = config_compat.patch_source(source_fixture)
assert 'layer_type == "qwen_sparse_attention"' in patched_fixture
assert '"full_attention"' in patched_fixture
assert config_compat.patch_source(patched_fixture) == patched_fixture

lm_head_source_fixture = """\
        self.lm_head = ParallelLMHead(
            config.vocab_size,
            config.hidden_size,
            prefix=maybe_prefix(prefix, "lm_head"),
        )
"""
lm_head_patched_fixture = lm_head_compat.patch_source(lm_head_source_fixture)
assert "quant_config=self.quant_config," in lm_head_patched_fixture
assert 'prefix="lm_head",' in lm_head_patched_fixture
assert lm_head_compat.patch_source(lm_head_patched_fixture) == lm_head_patched_fixture

mtp_lm_head_source_fixture = """\
                self.lm_head = ParallelLMHead(
                    config.vocab_size,
                    config.hidden_size,
                    prefix=maybe_prefix(prefix, "lm_head"),
                )
"""
mtp_lm_head_patched_fixture = lm_head_compat.patch_source(mtp_lm_head_source_fixture)
assert "                    quant_config=self.quant_config," in mtp_lm_head_patched_fixture
assert '                    prefix="lm_head",' in mtp_lm_head_patched_fixture

parallel_lm_head_fixture = """\
        self.quant_config = quant_config
        if bias:
            self._register_bias()
"""
parallel_lm_head_patched = lm_head_attrs.patch_source(parallel_lm_head_fixture)
assert "self.output_partition_sizes = [self.num_embeddings_per_partition]" in parallel_lm_head_patched
assert "self.has_bias = bias" in parallel_lm_head_patched
assert lm_head_attrs.patch_source(parallel_lm_head_patched) == parallel_lm_head_patched

fp8 = m._resolve_ple_dtype("F8_E4M3")
assert fp8.torch_dtype == torch.float8_e4m3fn
assert fp8.itemsize == 1 and fp8.needs_scale is True
assert m._row_bytes(160, fp8) == 160

bf16 = m._resolve_ple_dtype("BF16")
assert bf16.torch_dtype == torch.bfloat16
assert bf16.itemsize == 2 and bf16.needs_scale is False
assert m._row_bytes(160, bf16) == 320

try:
    m._resolve_ple_dtype("F32")
    raise AssertionError("unsupported PLE dtype must fail")
except ValueError as exc:
    assert "unsupported PLE shard dtype" in str(exc)

assert m._read_required_scale("BF16", None) is None
try:
    m._read_required_scale("F8_E4M3", None)
    raise AssertionError("FP8 PLE without a scale must fail")
except RuntimeError as exc:
    assert "FP8 shards without ngram_embedding.weight_scale" in str(exc)

ROWS, COLS, PARTS = 100_000, 160, 8
shard_size = -(-ROWS // PARTS)
rng = np.random.default_rng(0)
table = rng.integers(0, 256, size=(ROWS, COLS), dtype=np.uint8)

tmp = tempfile.mkdtemp()
# write shards into 2 safetensors files (4 shards each) with a dummy tensor first,
# so data offsets are non-trivial
file_of = {}
for fi in range(2):
    tensors = {"dummy.weight": np.arange(37, dtype=np.float32).tobytes()}
    header = {"dummy.weight": {"dtype": "F32", "shape": [37], "data_offsets": [0, 37 * 4]}}
    off = 37 * 4
    for si in range(fi * 4, fi * 4 + 4):
        rows = table[si * shard_size : (si + 1) * shard_size]
        name = f"model.language_model.layers.1.ple.ple_embedding.ngram_embedding.shard_{si}.weight"
        header[name] = {"dtype": "F8_E4M3", "shape": list(rows.shape), "data_offsets": [off, off + rows.nbytes]}
        tensors[name] = rows.tobytes()
        off += rows.nbytes
        file_of[name] = f"model-plefp8-0000{fi}.safetensors"
    if fi == 1:
        name = "model.language_model.layers.1.ple.ple_embedding.ngram_embedding.weight_scale"
        header[name] = {"dtype": "F32", "shape": [], "data_offsets": [off, off + 4]}
        tensors[name] = struct.pack("<f", 0.03125)
        off += 4
        file_of[name] = f"model-plefp8-0000{fi}.safetensors"
    hb = json.dumps(header).encode()
    with open(os.path.join(tmp, f"model-plefp8-0000{fi}.safetensors"), "wb") as f:
        f.write(struct.pack("<Q", len(hb)))
        f.write(hb)
        for name in header:
            f.write(tensors[name])
with open(os.path.join(tmp, "model.safetensors.index.json"), "w") as f:
    json.dump({"weight_map": file_of}, f)

shards, dtype_str, scale_entry = m._find_shards(tmp, 1)
cols = shards.pop("__cols__")
assert dtype_str == "F8_E4M3" and cols == COLS, (dtype_str, cols)
assert len(shards) == PARTS, len(shards)
assert abs(float(m._read_scale(scale_entry)) - 0.03125) < 1e-9
for idx, (_p, _o, rows) in shards.items():
    assert rows == max(0, min(shard_size, ROWS - idx * shard_size)), (idx, rows)

t = m.MmapPleTable(shards, shard_size, cols, torch.float8_e4m3fn, workers=8, chunk=512)
assert t.rows_total == ROWS

for n in (1, 16, 5000, 131_072):
    ids = rng.integers(0, ROWS, size=n, dtype=np.int64)
    ids[: n // 3] = ids[0]  # lots of duplicates, like real n-grams
    t0 = time.perf_counter()
    got = t.gather(ids)
    dt = time.perf_counter() - t0
    ref = table[ids]
    assert got.shape == (n, COLS) and got.dtype == np.uint8
    assert np.array_equal(got, ref), f"mismatch for n={n}"
    print(f"gather n={n:>7}: OK in {dt*1e3:7.2f} ms")

# torch view path used by the placeholder
emb = m._MmapNgramEmbedding(ROWS, COLS)
emb.table = t
ids_t = torch.from_numpy(rng.integers(0, ROWS, size=(300, 16), dtype=np.int64))
out = emb(ids_t)
assert out.shape == (300, 16, COLS) and out.dtype == torch.float8_e4m3fn
assert np.array_equal(out.view(torch.uint8).numpy().reshape(-1, COLS), table[ids_t.numpy().reshape(-1)])
print("placeholder forward: OK (fp8 view, shape", tuple(out.shape), ")")

# BF16 checkpoint path: exact values, a partial final shard, cross-shard ids,
# duplicates, and a non-trivial safetensors data offset.
BF_ROWS, BF_COLS, BF_PARTS = 1_031, 7, 8
bf_shard_size = -(-BF_ROWS // BF_PARTS)
bf_ref = (
    torch.arange(BF_ROWS * BF_COLS, dtype=torch.float32)
    .remainder(997)
    .div(31)
    .to(torch.bfloat16)
    .reshape(BF_ROWS, BF_COLS)
)
bf_u8 = bf_ref.view(torch.uint8).numpy().reshape(BF_ROWS, BF_COLS * 2)
bf_tmp = tempfile.mkdtemp()
bf_file_of = {}
for fi in range(2):
    tensors = {"dummy.weight": np.arange(11, dtype=np.float32).tobytes()}
    header = {
        "dummy.weight": {
            "dtype": "F32",
            "shape": [11],
            "data_offsets": [0, 11 * 4],
        }
    }
    off = 11 * 4
    for si in range(fi * 4, fi * 4 + 4):
        rows = bf_u8[si * bf_shard_size : (si + 1) * bf_shard_size]
        name = f"model.language_model.layers.1.ple.ple_embedding.ngram_embedding.shard_{si}.weight"
        header[name] = {
            "dtype": "BF16",
            "shape": [len(rows), BF_COLS],
            "data_offsets": [off, off + rows.nbytes],
        }
        tensors[name] = rows.tobytes()
        off += rows.nbytes
        bf_file_of[name] = f"model-plebf16-{fi}.safetensors"
    hb = json.dumps(header).encode()
    with open(os.path.join(bf_tmp, f"model-plebf16-{fi}.safetensors"), "wb") as f:
        f.write(struct.pack("<Q", len(hb)))
        f.write(hb)
        for name in header:
            f.write(tensors[name])
with open(os.path.join(bf_tmp, "model.safetensors.index.json"), "w") as f:
    json.dump({"weight_map": bf_file_of}, f)

bf_shards, bf_dtype, bf_scale = m._find_shards(bf_tmp, 1)
bf_cols = bf_shards.pop("__cols__")
assert bf_dtype == "BF16" and bf_scale is None and bf_cols == BF_COLS
bf_table = m._open_ple_table(
    bf_shards,
    bf_shard_size,
    bf_cols,
    bf_dtype,
    workers=4,
    chunk=32,
)
bf_ids_np = np.array(
    [0, bf_shard_size - 1, bf_shard_size, BF_ROWS - 1, 3, 3],
    dtype=np.int64,
)
bf_emb = m._MmapNgramEmbedding(BF_ROWS, BF_COLS)
bf_emb.table = bf_table
bf_got = bf_emb(torch.from_numpy(bf_ids_np))
assert bf_got.dtype == torch.bfloat16
assert torch.equal(bf_got.cpu(), bf_ref[torch.from_numpy(bf_ids_np)])
assert bf_table.row_bytes == BF_COLS * 2
print("placeholder forward: OK (bf16 exact values, shape", tuple(bf_got.shape), ")")

width_tmp = tempfile.mkdtemp()
width_names = [
    "model.language_model.layers.1.ple.ple_embedding.ngram_embedding.shard_0.weight",
    "model.language_model.layers.1.ple.ple_embedding.ngram_embedding.shard_1.weight",
]
width_header = {
    width_names[0]: {"dtype": "BF16", "shape": [1, 7], "data_offsets": [0, 14]},
    width_names[1]: {"dtype": "BF16", "shape": [1, 8], "data_offsets": [14, 30]},
}
width_hb = json.dumps(width_header).encode()
width_file = os.path.join(width_tmp, "model-width.safetensors")
with open(width_file, "wb") as f:
    f.write(struct.pack("<Q", len(width_hb)))
    f.write(width_hb)
    f.write(bytes(30))
with open(os.path.join(width_tmp, "model.safetensors.index.json"), "w") as f:
    json.dump({"weight_map": {name: "model-width.safetensors" for name in width_names}}, f)
try:
    m._find_shards(width_tmp, 1)
    raise AssertionError("mixed PLE shard widths must fail")
except ValueError as exc:
    assert "mixed widths" in str(exc)

layout_shards = {
    0: ("a", 0, 9),
    1: ("b", 0, 9),
    2: ("c", 0, 7),
}
assert m._validate_shard_layout(layout_shards, parts=3, vocab=25) == 9

try:
    m._validate_shard_layout({0: layout_shards[0], 2: layout_shards[2]}, parts=3, vocab=25)
    raise AssertionError("a missing middle PLE shard must fail")
except RuntimeError as exc:
    assert "PLE shard indices" in str(exc) and "[0, 2]" in str(exc)

bad_rows = dict(layout_shards)
bad_rows[2] = ("c", 0, 6)
try:
    m._validate_shard_layout(bad_rows, parts=3, vocab=25)
    raise AssertionError("a PLE shard with the wrong row count must fail")
except RuntimeError as exc:
    assert "PLE shard 2 has 6 rows, expected 7" in str(exc)

# zeros path (no table)
emb2 = m._MmapNgramEmbedding(ROWS, COLS)
z = emb2(ids_t)
assert z.shape == (300, 16, COLS) and float(z.abs().sum()) == 0.0
print("zeros path: OK")

# out-of-range must raise, not corrupt
try:
    t.gather(np.array([ROWS + 5], dtype=np.int64))
    raise SystemExit("expected IndexError")
except IndexError:
    print("out-of-range: raises IndexError OK")

t.prewarm()
print("prewarm: OK")
print("ALL OK")