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
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
|
#!/usr/bin/env bash
set -uo pipefail
if [[ -n "${ZSH_VERSION:-}" ]]; then
ROOT="$(cd -- "$(dirname -- "${(%):-%x}")" && pwd -P)"
else
ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
fi
export UV_CACHE_DIR="${UV_CACHE_DIR:-$ROOT/.venv/uv-cache}"
RED=""
GREEN=""
YELLOW=""
NORMAL=""
MAXWIDTH=0
if [[ -t 2 ]] && command -v tput >/dev/null 2>&1; then
RED="$(tput setaf 1 2>/dev/null || true)"
GREEN="$(tput setaf 2 2>/dev/null || true)"
YELLOW="$(tput setaf 3 2>/dev/null || true)"
NORMAL="$(tput sgr0 2>/dev/null || true)"
MAXWIDTH="$(tput cols 2>/dev/null || printf '0')"
if (( MAXWIDTH > 2 )); then
MAXWIDTH=$((MAXWIDTH - 2))
else
MAXWIDTH=0
fi
fi
info() { printf '%sINFO: %s%s\n' "$GREEN" "$*" "$NORMAL" >&2; }
warn() { printf '%sWARN: %s%s\n' "$YELLOW" "$*" "$NORMAL" >&2; }
err() { printf '%sERROR: %s%s\n' "$RED" "$*" "$NORMAL" >&2; }
_ALL_CMDS="help checks check-code-quality check-kernel-golden check-batch-parity \
check-context-batch-parity check-late-interaction check-cuda-fast-gemm-parity \
check-embedding-parity check-mlx-model-limits check-mlx-quantized-parity \
check-precision-drift check-reference-parity check-bert-parity check-e5-parity \
check-qwen3-parity check-tokenizer-parity check-wordpiece-parity check-model-matrix \
check-server-api check-cli check-workspace-api check-perplexity-sdk-compat \
check-openai-sdk-compat bench-contextual-api bench-late-rerank bench-late-api \
bench-matrix bench-qwen-tokenizer bench-server-api"
if [[ -n "${ZSH_VERSION:-}" ]]; then
_all_complete_zsh() {
# shellcheck disable=SC2296
local -a cmds=("${(ps: :)_ALL_CMDS}")
_arguments "1: :(${cmds[*]})"
}
compdef _all_complete_zsh ./all.bash all.bash all 2>/dev/null || true
else
_all_complete() {
local cur
cur="${COMP_WORDS[COMP_CWORD]}"
# shellcheck disable=SC2207
COMPREPLY=( $(compgen -W "$_ALL_CMDS" -- "$cur") )
}
complete -F _all_complete ./all.bash
complete -F _all_complete all.bash
complete -F _all_complete all
fi
run() {
local rendered
printf -v rendered '%q ' "$@"
rendered="${rendered% }"
if (( MAXWIDTH > 0 && ${#rendered} > MAXWIDTH )); then
rendered="${rendered:0:$((MAXWIDTH - 3))}..."
fi
info "\$ $rendered"
(cd "$ROOT" && "$@")
}
run_uv() {
run uv run "$@"
}
# Run a Python check that needs the pinned SentenceTransformers reference stack
# (torch, sentence-transformers); the pins live in requirements-reference.txt.
run_reference() {
run uv run --with-requirements requirements-reference.txt "$@"
}
check_code_quality() {
run ./devtools/check_code_quality.sh "$@"
}
# The comprehensive local suite, folded in from the old verify_builds.sh so
# all.bash is the single entry point. It builds, runs the hermetic C tests and
# the math-kernel goldens and the hermetic tokenizer checks (all model-free),
# then the CPU model checks and - on macOS - the MLX model checks, and finishes
# with the compiler/analyzer/tidy gate. Model-dependent steps run only when
# --model-dir is given. CUDA validation stays a separate RunPod step.
run_checks() {
local model_dir="" context_model_dir="" late_model_dir=""
local batch_size=4 context_runs=0
local long=0 skip_build=0 skip_hermetic=0 skip_mlx=0
local skip_code_quality=0 skip_tokenizer=0 skip_mlx_quantized=0
local platform_skip_mlx=0
while (($#)); do
case "$1" in
--model-dir) model_dir="$2"; shift 2 ;;
--model-dir=*) model_dir="${1#*=}"; shift ;;
--context-model-dir) context_model_dir="$2"; shift 2 ;;
--context-model-dir=*) context_model_dir="${1#*=}"; shift ;;
--late-model-dir) late_model_dir="$2"; shift 2 ;;
--late-model-dir=*) late_model_dir="${1#*=}"; shift ;;
--batch-size) batch_size="$2"; shift 2 ;;
--batch-size=*) batch_size="${1#*=}"; shift ;;
--context-runs) context_runs="$2"; shift 2 ;;
--context-runs=*) context_runs="${1#*=}"; shift ;;
--long) long=1; shift ;;
--skip-build) skip_build=1; shift ;;
--skip-hermetic) skip_hermetic=1; shift ;;
--skip-mlx) skip_mlx=1; shift ;;
--skip-code-quality) skip_code_quality=1; shift ;;
--skip-tokenizer) skip_tokenizer=1; shift ;;
--skip-mlx-quantized) skip_mlx_quantized=1; shift ;;
-h|--help)
cat <<'USAGE'
Usage: ./all.bash checks [options]
--model-dir DIR model for the workspace/batch parity checks
--context-model-dir DIR contextual model (enables contextual parity)
--late-model-dir DIR late-interaction model (enables late checks)
--batch-size N batch size for the parity checks (default 4)
--context-runs N timed contextual batches per backend
--long also run the long CPU attention parity check
--skip-build skip make cpu / make gpu
--skip-hermetic skip the model-free make test suite
--skip-mlx stop after the CPU build and CPU checks
--skip-tokenizer skip the hermetic tokenizer parity checks
--skip-mlx-quantized skip the MLX Q8 parity check
--skip-code-quality skip the compiler/analyzer/tidy gate
USAGE
return 0 ;;
*) err "checks: unknown option: $1"; return 2 ;;
esac
done
if [[ "$(uname -s)" != "Darwin" && $skip_mlx -eq 0 ]]; then
warn "MLX checks require macOS; skipping the MLX phase on $(uname -s)"
skip_mlx=1
platform_skip_mlx=1
fi
# Any failing step aborts the suite with a non-zero exit, as the old script did.
_step() { "$@" || { err "checks: step failed -> $*"; exit 1; }; }
local long_args=()
(( long )) && long_args=(--long)
# Model-free: build, hermetic C tests, kernel goldens, hermetic tokenizers.
(( skip_hermetic )) || _step run make test
(( skip_build )) || _step run make cpu
_step check_kernel_golden
if (( ! skip_tokenizer )); then
_step check_tokenizer_parity
_step check_wordpiece_parity
fi
# CPU model checks (need a model directory).
if [[ -n "$model_dir" ]]; then
_step check_workspace_api --model-dir "$model_dir"
_step check_batch_parity --model-dir "$model_dir" --backend cpu \
--batch-size "$batch_size" "${long_args[@]}"
if [[ -n "$context_model_dir" ]]; then
_step check_contextual_batch_parity --model-dir "$context_model_dir" \
--backend cpu --runs "$context_runs"
fi
if [[ -n "$late_model_dir" ]]; then
_step check_late_interaction --model-dir "$late_model_dir" --backend cpu
fi
else
warn "checks: no --model-dir; skipping CPU model checks (workspace, batch/contextual/late parity)"
fi
# MLX phase (macOS only): rebuild as the GPU backend, rerun the model checks.
if (( ! skip_mlx )); then
(( skip_build )) || _step run make gpu
if [[ -n "$model_dir" ]]; then
_step check_batch_parity --model-dir "$model_dir" --backend mlx \
--batch-size "$batch_size"
_step check_mlx_model_limits --model-dir "$model_dir"
(( skip_mlx_quantized )) || _step check_mlx_quantized_parity \
--model-dir "$model_dir" --batch-size "$batch_size"
if [[ -n "$context_model_dir" ]]; then
_step check_contextual_batch_parity --model-dir "$context_model_dir" \
--backend mlx --runs "$context_runs"
fi
if [[ -n "$late_model_dir" ]]; then
_step check_late_interaction --model-dir "$late_model_dir" --backend mlx
fi
fi
fi
(( skip_code_quality )) || _step check_code_quality
info "checks: all green"
}
check_kernel_golden() {
# Build and run the math-kernel golden tests straight from the test Makefile.
# test_kernels exercises the BLAS-built kernels; test_kernels_generic the
# scalar/SIMD path with USE_BLAS off. Extra args (e.g. --blas, kept for
# backward compatibility) are accepted but ignored; both variants always run.
run make -C tests test_kernels test_kernels_generic
run ./tests/test_kernels
run ./tests/test_kernels_generic
}
check_batch_parity() {
run_uv ./tests/check_batch_parity.py "$@"
}
check_contextual_batch_parity() {
run_uv ./tests/check_contextual_batch_parity.py "$@"
}
check_late_interaction() {
run_uv ./tests/check_late_interaction.py "$@"
}
check_cuda_fast_gemm_parity() {
if [[ "$(uname -s)" != "Linux" ]]; then
warn "CUDA fast GEMM parity requires Linux; skipping on $(uname -s)"
return 0
fi
if ! command -v nvidia-smi >/dev/null 2>&1; then
warn "CUDA fast GEMM parity requires an NVIDIA GPU; nvidia-smi not found"
return 0
fi
run_uv ./tests/check_cuda_fast_gemm_parity.py "$@"
}
check_embedding_parity() {
run_uv ./devtools/compare_model_embeddings.py "$@"
}
check_mlx_model_limits() {
if [[ "$(uname -s)" != "Darwin" ]]; then
warn "MLX metadata limits require macOS; skipping on $(uname -s)"
return 0
fi
run_uv ./tests/check_mlx_model_limits.py "$@"
}
check_mlx_quantized_parity() {
if [[ "$(uname -s)" != "Darwin" ]]; then
warn "MLX quantized parity requires macOS; skipping on $(uname -s)"
return 0
fi
run_uv ./tests/check_mlx_quantized_parity.py "$@"
}
check_quant_precision_drift() {
run_reference ./devtools/quant_precision_drift.py "$@"
}
check_reference_parity() {
run_reference ./tests/check_reference_parity.py "$@"
}
check_bert_parity() {
run_reference ./tests/check_bert_parity.py "$@"
}
check_e5_parity() {
run_reference ./tests/check_e5_parity.py "$@"
}
check_qwen3_parity() {
run uv run --with sentence-transformers --with torch --with numpy \
./tests/check_qwen3_parity.py "$@"
}
check_tokenizer_parity() {
run_uv ./tests/check_tokenizer_parity.py "$@"
}
check_wordpiece_parity() {
run_uv ./tests/check_wordpiece_parity.py "$@"
}
check_model_matrix() {
run_uv ./tests/check_model_matrix.py "$@"
}
check_perplexity_sdk_compat() {
run uv run --with perplexityai --with numpy \
./tests/check_perplexity_sdk_compat.py "$@"
}
check_openai_sdk_compat() {
run uv run --with openai ./tests/check_openai_sdk_compat.py "$@"
}
check_server_api() {
run_uv ./tests/check_server_api.py "$@"
}
check_cli() {
run_uv ./tests/check_cli.py "$@"
}
check_workspace_api() {
run_uv ./tests/check_workspace_api.py "$@"
}
bench_contextual_api() {
run_uv ./bench/bench_contextual_api.py "$@"
}
bench_late_rerank() {
run_uv ./bench/bench_late_rerank.py "$@"
}
bench_late_api() {
run_uv ./bench/bench_late_api.py "$@"
}
bench_matrix() {
run_uv ./bench/bench_matrix.py "$@"
}
bench_qwen_tokenizer() {
run_uv ./bench/bench_qwen_tokenizer.py "$@"
}
bench_server_api() {
run_uv ./bench/bench_server_api.py "$@"
}
usage() {
cat <<'EOF'
███████╗███████╗██╗ ██╗██████╗
██╔════╝██╔════╝██║ ██║██╔══██╗
█████╗ █████╗ ██║ █╗ ██║██║ ██║
██╔══╝ ██╔══╝ ██║███╗██║██║ ██║
██║ ██║ ╚███╔███╔╝██████╔╝
╚═╝ ╚═╝ ╚══╝╚══╝ ╚═════╝
Usage: ./all.bash COMMAND [ARGS...]
Aggregate: checks full local suite: build + hermetic tests + CPU/MLX
model checks + code quality. Pass --model-dir DIR to
add the model checks; `checks --help` for options.
┏━━[ checks ]━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ ┃
┃ check-code-quality Run clang builds, analysis ┃
┃ check-kernel-golden Run math-kernel golden tests ┃
┃ check-batch-parity Compare singleton vs batched ┃
┃ check-context-batch-parity Compare contextual execution ┃
┃ check-cuda-fast-gemm-parity Gate CUDA reduced-precision ┃
┃ check-embedding-parity Compare embeddings ┃
┃ check-late-interaction Check late-interaction vectors ┃
┃ check-mlx-model-limits Check MLX metadata rejection ┃
┃ check-mlx-quantized-parity Compare quantized embeddings ┃
┃ check-precision-drift Compare low-precision drift ┃
┃ check-reference-parity Compare server vs Python ref ┃
┃ check-bert-parity Compare BERT embeddings ┃
┃ check-e5-parity Compare XLM-R/E5 embeddings ┃
┃ check-qwen3-parity Compare Qwen3 embeddings ┃
┃ check-tokenizer-parity Compare BPE tokenizer ┃
┃ check-wordpiece-parity Compare WordPiece tokenizer ┃
┃ check-model-matrix Smoke-check model variants ┃
┃ check-server-api Verify running HTTP server ┃
┃ check-cli Drive CLI through modes/flags ┃
┃ check-workspace-api Check model/workspace API ┃
┃ check-perplexity-sdk-compat Verify via Perplexity SDK ┃
┃ check-openai-sdk-compat Verify via OpenAI SDK ┃
┃ ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
┏━━[ benchmarks ]━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ ┃
┃ bench-contextual-api Benchmark contextual inference ┃
┃ bench-late-rerank Benchmark late-interaction ┃
┃ bench-late-api Benchmark HTTP reranking ┃
┃ bench-matrix Run batch-aware stdin matrix ┃
┃ bench-qwen-tokenizer Benchmark tokenizer encode ┃
┃ bench-server-api Benchmark concurrent HTTP ┃
┃ ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛
EOF
}
main() {
local command="${1:-help}"
if (($#)); then
shift
fi
case "$command" in
help|-h|--help) usage ;;
checks) run_checks "$@" ;;
check-code-quality) check_code_quality "$@" ;;
check-kernel-golden) check_kernel_golden "$@" ;;
check-batch-parity) check_batch_parity "$@" ;;
check-context-batch-parity) check_contextual_batch_parity "$@" ;;
check-late-interaction) check_late_interaction "$@" ;;
check-cuda-fast-gemm-parity) check_cuda_fast_gemm_parity "$@" ;;
check-embedding-parity) check_embedding_parity "$@" ;;
check-mlx-model-limits) check_mlx_model_limits "$@" ;;
check-mlx-quantized-parity) check_mlx_quantized_parity "$@" ;;
check-precision-drift) check_quant_precision_drift "$@" ;;
check-reference-parity) check_reference_parity "$@" ;;
check-bert-parity) check_bert_parity "$@" ;;
check-e5-parity) check_e5_parity "$@" ;;
check-qwen3-parity) check_qwen3_parity "$@" ;;
check-tokenizer-parity) check_tokenizer_parity "$@" ;;
check-wordpiece-parity) check_wordpiece_parity "$@" ;;
check-model-matrix) check_model_matrix "$@" ;;
check-perplexity-sdk-compat) check_perplexity_sdk_compat "$@" ;;
check-openai-sdk-compat) check_openai_sdk_compat "$@" ;;
check-server-api) check_server_api "$@" ;;
check-cli) check_cli "$@" ;;
check-workspace-api) check_workspace_api "$@" ;;
bench-contextual-api) bench_contextual_api "$@" ;;
bench-late-rerank) bench_late_rerank "$@" ;;
bench-late-api) bench_late_api "$@" ;;
bench-matrix) bench_matrix "$@" ;;
bench-qwen-tokenizer) bench_qwen_tokenizer "$@" ;;
bench-server-api) bench_server_api "$@" ;;
*)
err "unknown command: $command"
usage >&2
return 2
;;
esac
}
if [[ -n "${ZSH_VERSION:-}" ]]; then
[[ "${ZSH_EVAL_CONTEXT:-}" == "toplevel" ]] && main "$@"
else
[[ "${BASH_SOURCE[0]}" == "$0" ]] && main "$@"
fi
|