File size: 9,044 Bytes
2f6d104
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Generative fusion-chain grammar: [+residual] -> {rms|layer}norm -> ×w(+b) -> epilogue.

The widest sweep of the reduction->epilogue region (where the compiler under-fuses and the 2B
wins). References are COMPOSED; teacher kernels are TEMPLATE-GENERATED (scalar-reduce +
whole-row variants). Everything is harness-filtered downstream — nothing here is trusted.
"""
from __future__ import annotations
import torch

_C = 0.7978845608028654  # sqrt(2/pi)

# epilogue: (torch fn on fp32 tensor, triton expression in fp32 var `n`)
ACTS = {
    "gelu":  (lambda t: 0.5 * t * (1.0 + torch.tanh(_C * (t + 0.044715 * t * t * t))),
              "(0.5 * n * (1.0 + (2.0 * tl.sigmoid(2.0 * (0.7978845608028654 * (n + 0.044715 * n * n * n))) - 1.0)))"),
    "silu":  (lambda t: t * torch.sigmoid(t), "(n * tl.sigmoid(n))"),
    "relu2": (lambda t: torch.relu(t) * torch.relu(t), "(tl.maximum(n, 0.0) * tl.maximum(n, 0.0))"),
    # --- expanded grammar (each torch fn EXACTLY matches its triton expr; no approximation) ---
    "tanh":    (lambda t: torch.tanh(t), "(2.0 * tl.sigmoid(2.0 * n) - 1.0)"),   # identity tanh(x)=2σ(2x)-1
    "sigmoid": (lambda t: torch.sigmoid(t), "tl.sigmoid(n)"),
    "relu":    (lambda t: torch.relu(t), "tl.maximum(n, 0.0)"),
    "square":  (lambda t: t * t, "(n * n)"),
    # --- 2c round 2: more real activations (each torch fn EXACTLY matches its triton expr) ---
    "abs":      (lambda t: torch.abs(t), "tl.abs(n)"),
    "softsign": (lambda t: t / (1.0 + torch.abs(t)), "(n / (1.0 + tl.abs(n)))"),
    "hardsigmoid": (lambda t: torch.clamp(t + 3.0, 0.0, 6.0) / 6.0,
                    "(tl.minimum(tl.maximum(n + 3.0, 0.0), 6.0) / 6.0)"),          # F.hardsigmoid
    "hardswish":   (lambda t: t * torch.clamp(t + 3.0, 0.0, 6.0) / 6.0,
                    "(n * tl.minimum(tl.maximum(n + 3.0, 0.0), 6.0) / 6.0)"),      # F.hardswish
    # --- V2 round 3: 8 more real, numerically-safe activations. Same exactness rule: the
    #     torch lambda IS the triton expression (tanh via the 2*sigmoid(2x)-1 identity;
    #     softplus uses F.softplus's threshold=20 guard so exp never overflows). -----------
    "leaky_relu": (lambda t: torch.where(t > 0, t, 0.01 * t),
                   "tl.where(n > 0.0, n, 0.01 * n)"),
    "relu6":      (lambda t: torch.clamp(t, 0.0, 6.0),
                   "tl.minimum(tl.maximum(n, 0.0), 6.0)"),
    "hardtanh":   (lambda t: torch.clamp(t, -1.0, 1.0),
                   "tl.minimum(tl.maximum(n, -1.0), 1.0)"),
    "elu":        (lambda t: torch.where(t > 0, t, torch.exp(torch.clamp(t, max=0.0)) - 1.0),
                   "tl.where(n > 0.0, n, tl.exp(tl.minimum(n, 0.0)) - 1.0)"),
    "selu":       (lambda t: 1.0507009873554805 * torch.where(
                       t > 0, t, 1.6732632423543772 * (torch.exp(torch.clamp(t, max=0.0)) - 1.0)),
                   "(1.0507009873554805 * tl.where(n > 0.0, n, "
                   "1.6732632423543772 * (tl.exp(tl.minimum(n, 0.0)) - 1.0)))"),
    "softplus":   (lambda t: torch.where(t > 20.0, t, torch.log(1.0 + torch.exp(torch.clamp(t, max=20.0)))),
                   "tl.where(n > 20.0, n, tl.log(1.0 + tl.exp(tl.minimum(n, 20.0))))"),
    "mish":       (lambda t: t * torch.tanh(torch.where(
                       t > 20.0, t, torch.log(1.0 + torch.exp(torch.clamp(t, max=20.0))))),
                   "(n * (2.0 * tl.sigmoid(2.0 * tl.where(n > 20.0, n, "
                   "tl.log(1.0 + tl.exp(tl.minimum(n, 20.0))))) - 1.0))"),
    "gelu_erf":   (lambda t: 0.5 * t * (1.0 + torch.erf(t * 0.7071067811865476)),
                   "(0.5 * n * (1.0 + tl.erf(n * 0.7071067811865476)))"),          # EXACT gelu
}
NORMS = ["rms", "layer"]
RESID = [False, True]
ACTNAMES = ["gelu", "silu", "relu2", "tanh", "sigmoid", "relu", "square",
            "abs", "softsign", "hardsigmoid", "hardswish",
            "leaky_relu", "relu6", "hardtanh", "elu", "selu", "softplus", "mish", "gelu_erf"]


def chain_name(norm, residual, act):
    return ("add_" if residual else "") + ("rmsnorm" if norm == "rms" else "layernorm") + "_" + act


def chain_kind(norm, residual):
    return ("add_" if residual else "") + ("rms" if norm == "rms" else "ln")   # -> input signature


def chain_reference(norm, residual, act, eps=None):
    eps = eps if eps is not None else (1e-6 if norm == "rms" else 1e-5)
    fn = ACTS[act][0]
    def ref(*args):
        if residual and norm == "rms":
            x, r, w = args; h = x.float() + r.float(); b = None
        elif residual:
            x, r, w, b = args; h = x.float() + r.float()
        elif norm == "rms":
            x, w = args; h = x.float(); b = None
        else:
            x, w, b = args; h = x.float()
        if norm == "rms":
            n = h * torch.rsqrt(h.pow(2).mean(-1, keepdim=True) + eps) * w.float()
        else:
            mu = h.mean(-1, keepdim=True); hc = h - mu
            n = hc * torch.rsqrt((hc * hc).mean(-1, keepdim=True) + eps) * w.float() + b.float()
        return fn(n).to(args[0].dtype)
    return ref


# ---- teacher-kernel templates -----------------------------------------------------------
def _kernel(norm, residual, act_expr, eps, variant):
    """variant: 'scalar' (loop+scalar accumulator) or 'whole' (single block per row)."""
    ptrs = "x_ptr, " + ("r_ptr, " if residual else "") + "w_ptr, " + ("b_ptr, " if norm == "layer" else "") + "y_ptr"
    sig = "x, " + ("residual, " if residual else "") + "w" + (", b" if norm == "layer" else "")
    launch = "x, " + ("residual, " if residual else "") + "w" + (", b" if norm == "layer" else "") + ", y"
    radv = " r_ptr += row * stride;" if residual else ""
    hload = ("tl.load(x_ptr + cols, mask=MM, other=0.0).to(tl.float32)"
             + (" + tl.load(r_ptr + cols, mask=MM, other=0.0).to(tl.float32)" if residual else ""))
    # bias load indent differs: scalar variant loads it INSIDE the apply for-loop (8 spaces),
    # whole-row loads it flat (4 spaces). Wrong indent -> IndentationError.
    bload8 = "        b = tl.load(b_ptr + cols, mask=MM, other=0.0).to(tl.float32)\n" if norm == "layer" else ""
    bload4 = "    b = tl.load(b_ptr + cols, mask=MM, other=0.0).to(tl.float32)\n" if norm == "layer" else ""
    if norm == "rms":
        normed = "h * rr * w"
    else:
        normed = "(h - mu) * rr * w + b"

    if variant == "scalar":
        if norm == "rms":
            reduce_block = f'''    s = 0.0
    for off in range(0, N, BLOCK):
        cols = off + tl.arange(0, BLOCK); MM = cols < N
        h = {hload}
        s += tl.sum(h * h)
    rr = tl.rsqrt(s / N + eps)'''
        else:
            reduce_block = f'''    s = 0.0
    for off in range(0, N, BLOCK):
        cols = off + tl.arange(0, BLOCK); MM = cols < N
        s += tl.sum({hload})
    mu = s / N
    v = 0.0
    for off in range(0, N, BLOCK):
        cols = off + tl.arange(0, BLOCK); MM = cols < N
        d = tl.where(MM, ({hload}) - mu, 0.0); v += tl.sum(d * d)
    rr = tl.rsqrt(v / N + eps)'''
        body = f'''@triton.jit
def _k({ptrs}, stride, N, eps, BLOCK: tl.constexpr):
    row = tl.program_id(0); x_ptr += row * stride;{radv} y_ptr += row * stride
{reduce_block}
    for off in range(0, N, BLOCK):
        cols = off + tl.arange(0, BLOCK); MM = cols < N
        h = {hload}
        w = tl.load(w_ptr + cols, mask=MM, other=0.0).to(tl.float32)
{bload8}        n = {normed}
        tl.store(y_ptr + cols, {act_expr}, mask=MM)
def run({sig}):
    M, N = x.shape; y = torch.empty_like(x)
    _k[(M,)]({launch}, x.stride(0), N, {eps}, BLOCK=1024)
    return y
'''
    else:  # whole-row single block
        if norm == "rms":
            stat = "    rr = tl.rsqrt(tl.sum(h * h) / N + eps)"
        else:
            stat = ("    mu = tl.sum(h) / N\n    hc = tl.where(MM, h - mu, 0.0)\n"
                    "    rr = tl.rsqrt(tl.sum(hc * hc) / N + eps)")
            normed = "hc * rr * w + b"
        body = f'''@triton.jit
def _k({ptrs}, stride, N, eps, BLOCK: tl.constexpr):
    row = tl.program_id(0); x_ptr += row * stride;{radv} y_ptr += row * stride
    cols = tl.arange(0, BLOCK); MM = cols < N
    h = {hload}
{stat}
    w = tl.load(w_ptr + cols, mask=MM, other=0.0).to(tl.float32)
{bload4}    n = {normed}
    tl.store(y_ptr + cols, {act_expr}, mask=MM)
def run({sig}):
    M, N = x.shape; y = torch.empty_like(x)
    _k[(M,)]({launch}, x.stride(0), N, {eps}, BLOCK=triton.next_power_of_2(N))
    return y
'''
    return body


def chain_structures(norm, residual, act):
    eps = 1e-6 if norm == "rms" else 1e-5
    expr = ACTS[act][1]
    return [_kernel(norm, residual, expr, eps, "scalar"), _kernel(norm, residual, expr, eps, "whole")]


def all_chains():
    """[(name, kind, reference_fn, [kernel_src, ...]), ...] for the full grammar."""
    out = []
    for norm in NORMS:
        for residual in RESID:
            for act in ACTNAMES:
                name = chain_name(norm, residual, act)
                out.append((name, chain_kind(norm, residual), chain_reference(norm, residual, act),
                            chain_structures(norm, residual, act)))
    return out