File size: 7,029 Bytes
7a102c1
 
 
 
 
16c436e
 
7a102c1
 
7011ac6
 
 
16c436e
 
7011ac6
 
7a102c1
 
 
7011ac6
7a102c1
 
 
16c436e
7011ac6
16c436e
 
 
 
 
 
 
 
 
 
7a102c1
 
 
 
 
 
 
7011ac6
 
7a102c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7011ac6
7a102c1
 
 
 
 
7011ac6
 
 
7a102c1
 
 
16c436e
 
 
7a102c1
 
 
 
 
 
 
 
 
 
 
 
16c436e
 
 
 
7011ac6
7a102c1
 
7011ac6
 
 
 
 
 
 
 
 
 
 
 
 
 
7a102c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""The Core AI backends of decide.py: d1's exported decision function on the GPU or the Neural Engine, run by the Swift
worker (swift/.build/release/d1worker), one warm process per backend, started on first use.

    backends = CoreAI(model, QUESTIONS)        # model: the mlx-vlm d1 model (its processor and readout are reused)
    backends.available()                       # {"coreai-gpu": None or why not, "coreai-ane": ...}
    backends.start("coreai-gpu")               # lazily, once: downloads the model the first time (3.1 GB), then specialises
                                               # (74 s on the first Neural Engine run ever), then seconds
    backends.decide("coreai-gpu", image)       # system_one's output for a PIL frame

The questions are an input of the exported graph: the prompt for QUESTIONS is built here at start-up (it needs the
input embedding table, embed_tokens_fp16.npy, 500 MB, downloaded with the first model) and any QUESTIONS whose prompt
fits the export's sizes (d1_common.T_MAX etc.) run on the same models.

The models come from https://huggingface.co/ivanfioravanti/d1-2B-gestures-CoreAI (Hugging Face cache, offline after the
first download); D1_COREAI_DIR=/path/to/folder with d1_gpu_prompt_fp16.aimodel / d1_ane_prompt_fp16.aimodel (and
embed_tokens_fp16.npy, else coreai/out's or the repo's) uses your own export instead.
"""
import json, os, subprocess, tempfile, time
import numpy as np
from d1_common import S_MAX, build_meta, embeddings, frame, pixels, prompt_inputs, prompt_problem, readout, uncrop_box

HERE = os.path.dirname(os.path.abspath(__file__))
WORKER = os.path.join(HERE, "swift", ".build", "release", "d1worker")
REPO = "ivanfioravanti/d1-2B-gestures-CoreAI"
MODELS = {"coreai-gpu": ("d1_gpu_prompt_fp16.aimodel", "gpu"), "coreai-ane": ("d1_ane_prompt_fp16.aimodel", "ane")}
LOCAL = os.environ.get("D1_COREAI_DIR")


def model_path(name):
    """The .aimodel folder: from D1_COREAI_DIR, else downloaded from the Hugging Face repo on first use (then cached)."""
    folder = MODELS[name][0]
    if LOCAL:
        return os.path.join(LOCAL, folder)
    from huggingface_hub import snapshot_download
    return os.path.join(snapshot_download(REPO, allow_patterns=[folder + "/*"]), folder)


class Unavailable(Exception):
    pass


class Worker:
    def __init__(self, path, compute, prompt_dir):
        self.p = subprocess.Popen([WORKER, "--model", path, "--compute", compute, "--prompt", prompt_dir], stdin=subprocess.PIPE,
                                  stdout=subprocess.PIPE, text=True, bufsize=1)
        self.ready = self.read()  # specialises (first time ever: minutes on the Neural Engine) and runs once
        self.pix = tempfile.NamedTemporaryFile(prefix="d1-%s-" % compute, suffix=".f32", delete=False).name

    def read(self):
        line = self.p.stdout.readline()
        if not line:
            raise RuntimeError("d1worker exited (code %s)" % self.p.poll())
        out = json.loads(line)
        if "error" in out:
            raise RuntimeError("d1worker: " + out["error"])
        return out

    def logits(self, x):
        x.astype(np.float32).tofile(self.pix)
        self.p.stdin.write(json.dumps({"pixels": self.pix}) + "\n")
        self.p.stdin.flush()
        return np.array(self.read()["logits"], dtype=np.float32).reshape(S_MAX, -1)


class CoreAI:
    def __init__(self, model, questions):
        self.engine, self.questions, self.workers = model.engine, questions, {}
        self.meta = build_meta(self.engine, questions, write=False)  # the prompt for these QUESTIONS
        self.problem = prompt_problem(self.meta)
        self.prompt_dir = None

    def available(self):
        out = {}
        for name, (folder, _) in MODELS.items():
            local = LOCAL and os.path.join(LOCAL, folder)
            out[name] = (self.problem or (None if not local or os.path.exists(local) else "missing %s (D1_COREAI_DIR)" % local)
                         or (None if os.path.exists(WORKER) else "missing the Swift worker: cd coreai/swift && swift build -c release"))
        return out

    def start(self, name):
        """The warm worker for a backend, started on first use; returns the seconds this call spent starting it."""
        why = self.available().get(name, "unknown backend")
        if why:
            raise Unavailable(why)
        w = self.workers.get(name)
        if w is not None and w.p.poll() is None:
            return 0.0
        t = time.perf_counter()
        try:
            path = model_path(name)
        except Exception as e:  # offline before the first download, or the repo unreachable
            raise Unavailable("could not download %s from huggingface.co/%s: %s" % (MODELS[name][0], REPO, e))
        self.workers[name] = Worker(path, MODELS[name][1], self.prompt())
        return time.perf_counter() - t

    def prompt(self):
        """The prompt inputs for QUESTIONS as the worker reads them, written once."""
        if self.prompt_dir is None:
            emb = np.load(os.path.join(LOCAL, "embed_tokens_fp16.npy"), mmap_mode="r") if LOCAL and os.path.exists(
                os.path.join(LOCAL, "embed_tokens_fp16.npy")) else embeddings()
            d = tempfile.mkdtemp(prefix="d1-prompt-")
            arrays = prompt_inputs(self.meta, emb)
            for k, v in arrays.items():
                v.tofile(os.path.join(d, k + ".f16"))
            with open(os.path.join(d, "shapes.json"), "w") as f:
                json.dump({k: list(v.shape) for k, v in arrays.items()}, f)
            self.prompt_dir = d
        return self.prompt_dir

    def decide(self, name, img):
        w = self.workers[name]
        try:
            logits = w.logits(pixels(self.engine, frame(img)))  # frame(): the 640x360 the export was traced at
        except Exception:
            self.workers.pop(name, None)  # a dead worker is restarted on the next request
            raise
        answers = readout(self.engine, self.questions, logits, self.meta)
        for a in answers.values():  # boxes come back in the crop's coordinates: map them onto the whole frame
            if a.get("type") == "box" and a.get("box"):
                a["box"] = uncrop_box(a["box"], *img.size)
        return {"answers": answers, "usage": {"input_tokens": len(self.meta["ids"]), "output_tokens": 0}}


if __name__ == "__main__":  # the crop arithmetic: python coreai_backend.py
    full = {"top": 0.0, "bottom": 1.0, "left": 0.0, "right": 1.0}
    assert uncrop_box(full, 640, 360) == full                                  # 16:9: unchanged
    assert uncrop_box(full, 400, 400) == {"top": 87 / 400, "bottom": 312 / 400, "left": 0.0, "right": 1.0}  # 400x225 band
    assert uncrop_box(full, 1000, 360) == {"top": 0.0, "bottom": 1.0, "left": 180 / 1000, "right": 820 / 1000}
    b = uncrop_box({"top": 0.5, "bottom": 0.5, "left": 0.25, "right": 0.75}, 640, 480)  # 640x360 band at y 60..420
    assert b == {"top": 240 / 480, "bottom": 240 / 480, "left": 0.25, "right": 0.75}, b
    print("uncrop_box ok")