File size: 6,344 Bytes
0f775e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Llama-shaped decoder with INT4 group-quantised weights (W4A16), shipped pre-packed.

Every 2-D weight arrives as a triple `(packed, scale, zero)`:

    packed : (out, in // 2) uint8   -- two 4-bit codes per byte, LOW nibble first
    scale  : (out, in // group) bf16
    zero   : (out, in // group) uint8 (values 0..15)

and its value is `(code - zero) * scale`, with `group` = `cfg["group"]` contiguous input elements
sharing one (scale, zero). This is the AWQ / GPTQ asymmetric layout.

The weights are quantised ONCE, here, and the reference dequantises exactly these bytes. That is not a
convenience: if the fixture were fp32 and the agent had to quantise, a *correct* int4 kernel would
disagree with the reference by the quantisation error rather than by its own error -- measured at 10x
the tolerance for the fp8 case.
"""
from model import HELPERS_CORE

QUANT = r'''

_INT4_GROUP = None                                   # set by make_weights / build_model from cfg


def _quantise(w, dt, group=128):
    """fp32 -> asymmetric int4 with per-group (scale, zero), packed two codes per byte."""
    if dt == "bf16":
        return w.to(torch.bfloat16)
    if dt != "int4":
        raise ValueError(dt)
    out, inn = w.shape
    g = w.view(out, inn // group, group)
    lo = g.amin(dim=-1, keepdim=True)
    hi = g.amax(dim=-1, keepdim=True)
    scale = ((hi - lo) / 15.0).clamp(min=1e-8)
    zero = torch.round(-lo / scale).clamp(0, 15)
    code = torch.round(g / scale + zero).clamp(0, 15).to(torch.uint8).view(out, inn)
    packed = (code[:, 0::2] | (code[:, 1::2] << 4)).contiguous()
    return (packed, scale.squeeze(-1).to(torch.bfloat16), zero.squeeze(-1).to(torch.uint8))


def _deq(w, group=128):
    """(packed, scale, zero) -> bf16. Plain bf16 weights pass through."""
    if not isinstance(w, tuple):
        return w
    packed, scale, zero = w
    out = packed.shape[0]
    lo = (packed & 0xF).to(torch.int16)
    hi = (packed >> 4).to(torch.int16)
    code = torch.stack([lo, hi], dim=-1).view(out, -1)            # interleave back to (out, in)
    inn = code.shape[1]
    code = code.view(out, inn // group, group).float()
    v = (code - zero.float().unsqueeze(-1)) * scale.float().unsqueeze(-1)
    return v.view(out, inn).to(torch.bfloat16)
'''

BODY = r'''
def make_weights(cfg, seed=0, device="cuda"):
    """Deterministic 1/sqrt(fan_in)-scaled weights, shipped ALREADY int4-quantised."""
    g = torch.Generator(device=device).manual_seed(seed)
    d, ffn, n_q, n_kv, hd = cfg["d"], cfg["ffn"], cfg["n_q"], cfg["n_kv"], cfg["hd"]
    dt, grp = cfg["wdtype"], cfg["group"]

    def rnd(*shape, fan_in):
        w = torch.randn(*shape, device=device, dtype=torch.float32, generator=g) / (fan_in ** 0.5)
        return _quantise(w, dt, grp)

    ones = lambda: torch.ones(d, device=device, dtype=torch.bfloat16)
    W = {"embed": rnd(cfg["vocab"], d, fan_in=d), "final_norm": ones(), "layers": []}
    for _ in range(cfg["layers"]):
        W["layers"].append(dict(
            in_norm=ones(), post_norm=ones(),
            q=rnd(n_q * hd, d, fan_in=d), k=rnd(n_kv * hd, d, fan_in=d),
            v=rnd(n_kv * hd, d, fan_in=d), o=rnd(d, n_q * hd, fan_in=n_q * hd),
            gate=rnd(ffn, d, fan_in=d), up=rnd(ffn, d, fan_in=d), down=rnd(d, ffn, fan_in=ffn)))
    return W


def make_kv(cfg, batch, prefill_len, max_seq, seed=0, device="cuda"):
    """KV cache already holding `prefill_len` tokens. Decode starts at pos = prefill_len."""
    g = torch.Generator(device=device).manual_seed(seed + 777)
    kv = []
    for _ in range(cfg["layers"]):
        k = torch.zeros(batch, cfg["n_kv"], max_seq, cfg["hd"], device=device, dtype=torch.bfloat16)
        v = torch.zeros_like(k)
        k[:, :, :prefill_len] = torch.randn(batch, cfg["n_kv"], prefill_len, cfg["hd"], device=device,
                                            dtype=torch.float32, generator=g).to(torch.bfloat16) * 0.5
        v[:, :, :prefill_len] = torch.randn(batch, cfg["n_kv"], prefill_len, cfg["hd"], device=device,
                                            dtype=torch.float32, generator=g).to(torch.bfloat16) * 0.5
        kv.append((k, v))
    return kv


def build_model(weights, kv_cache, cfg, max_seq_len):
    """UNTIMED setup. Dequantises ONCE here rather than per step -- dequantising inside every step
    allocates GBs per call, which perturbs the caching allocator enough that cuBLAS picks different
    GEMV algorithms run-to-run and two bit-identical implementations drift apart."""
    cos, sin = _rope_cache(cfg, max_seq_len, weights["final_norm"].device)
    grp = cfg["group"]
    W = {"embed": _deq(weights["embed"], grp), "final_norm": weights["final_norm"],
         "layers": [{k: (v if k.endswith("norm") else _deq(v, grp)) for k, v in L.items()}
                    for L in weights["layers"]]}
    return {"W": W, "kv": kv_cache, "cfg": cfg, "cos": cos, "sin": sin}


@torch.no_grad()
def decode_step(handle, token_ids, pos):
    """One decode step for every sequence in the batch. Appends this position's K/V into the cache.

    token_ids: (B,) int64      pos: int, the absolute position being written
    returns:   (B, vocab) logits
    """
    W, kv, cfg = handle["W"], handle["kv"], handle["cfg"]
    cos, sin = handle["cos"], handle["sin"]
    B = token_ids.shape[0]
    n_q, n_kv, hd = cfg["n_q"], cfg["n_kv"], cfg["hd"]
    rep = n_q // n_kv

    x = W["embed"][token_ids]
    for li, L in enumerate(W["layers"]):
        h = _rms_norm(x, L["in_norm"], cfg["eps"])
        q = (h @ L["q"].T).view(B, n_q, 1, hd)
        k = (h @ L["k"].T).view(B, n_kv, 1, hd)
        v = (h @ L["v"].T).view(B, n_kv, 1, hd)
        q = _apply_rope(q, cos, sin, pos)
        k = _apply_rope(k, cos, sin, pos)
        kc, vc = kv[li]
        kc[:, :, pos:pos + 1] = k
        vc[:, :, pos:pos + 1] = v
        kk = kc[:, :, :pos + 1].repeat_interleave(rep, dim=1)
        vv = vc[:, :, :pos + 1].repeat_interleave(rep, dim=1)
        att = F.scaled_dot_product_attention(q, kk, vv)
        x = x + (att.reshape(B, n_q * hd) @ L["o"].T)
        h = _rms_norm(x, L["post_norm"], cfg["eps"])
        x = x + ((F.silu(h @ L["gate"].T) * (h @ L["up"].T)) @ L["down"].T)
    x = _rms_norm(x, W["final_norm"], cfg["eps"])
    return x @ W["embed"].T                      # tied lm_head
'''

MODEL_SRC = HELPERS_CORE + QUANT + BODY