LlamaTov · Bit-Perfect Dispatch

🕯️ LlamaTov · Bit-Perfect Dispatch

Two-axis 0-ULP congruence matrix — RUNTIME (compute vs oracle) × MIGRATION (source-preservation swipl→swilgt). Generated 2026-09-02T22:06:02Z 10d 3h ago · auto-refresh 10s

vs stock (runtime)

21/ 22 runtime 0-ULP bit-identical

1within tolerance (open cells)

0diverged beyond tolerance

vs truth (accuracy)

21matched (vs truth, entailed by bit-identity)

1IMPROVED vs truth (measured, better than stock)

migration (swipl→swilgt)

21/ 21 migration source-identical (swipl→swilgt)

measurement

493,395floats compared (population across all kernels)

6rows under-exercised (< 1000 floats)

21 / 22 runtime 0-ULP (95%)

Kernel	Shape	Runtime	max ULP	Floats	Oracle	Backend/Dev	Accuracy (vs truth)	Migration
sgemm_cpu	64x64	BIT_IDENTICAL	0	4,096	torch.matmul	cpu float32 cpu	MATCHED	-
sgemm_cpu	256x256	BIT_IDENTICAL	0	65,536	torch.matmul	cpu float32 cpu	MATCHED	-
sgemm_cpu	512x512	BIT_IDENTICAL	0	262,144	torch.matmul	cpu float32 cpu	MATCHED	-
relu_cpu	10000	BIT_IDENTICAL	0	10,000	torch.nn.functional.relu	cpu float32 cpu	MATCHED	-
silu_cpu	10000	BIT_IDENTICAL	0	10,000	torch.nn.functional.silu	cpu float32 cpu	MATCHED	-
mish_cpu	10000	BIT_IDENTICAL	0	10,000	torch.nn.functional.mish	cpu float32 cpu	MATCHED	-
fused_mm_bias_relu_cpu	256x256	BIT_IDENTICAL	0	65,536	torch.relu(x@W+b)	cpu float32 cpu	MATCHED	-
conv2d_cpu	1x3x16x16	BIT_IDENTICAL	0	2,048	torch.nn.functional.conv2d	cpu float32 cpu	MATCHED	-
upsample_cpu	1x8x4x4	BIT_IDENTICAL	0	512under-exercised	torch.nn.functional.interpolate	cpu float32 cpu	MATCHED	-
sigmoid_cpu	10000	BIT_IDENTICAL	0	10,000	torch.nn.functional.sigmoid	cpu float32 cpu	MATCHED	-
tanh_cpu	10000	BIT_IDENTICAL	0	10,000	torch.nn.functional.tanh	cpu float32 cpu	MATCHED	-
gelu_cpu	10000	PASS_ABS_TOLERANCE	127951	7,682 / 10,000	torch.nn.functional.gelu	cpu float32 cpu	IMPROVED \|err\| 1.09e-08/3.86e-08 (3.5× better) · rounded 6,211/2,336	-
neg_cpu	10000	BIT_IDENTICAL	0	10,000	torch.neg	cpu float32 cpu	MATCHED	-
abs_cpu	10000	BIT_IDENTICAL	0	10,000	torch.abs	cpu float32 cpu	MATCHED	-
exp_cpu	10000	BIT_IDENTICAL	0	10,000	torch.exp	cpu float32 cpu	MATCHED	-
reduce_sum_cpu	1024	BIT_IDENTICAL	0	1under-exercised	torch.sum	cpu float32 cpu	MATCHED	-
reduce_mean_cpu	1024	BIT_IDENTICAL	0	1under-exercised	torch.mean	cpu float32 cpu	MATCHED	-
reduce_max_cpu	1024	BIT_IDENTICAL	0	1under-exercised	torch.max	cpu float32 cpu	MATCHED	-
softmax_cpu	32x64	BIT_IDENTICAL	0	2,048	torch.nn.functional.softmax	cpu float32 cpu	MATCHED	-
layernorm_cpu	8x128	BIT_IDENTICAL	0	1,024	torch.nn.functional.layer_norm	cpu float32 cpu	MATCHED	-
maxpool2d_cpu	1x3x16x16	BIT_IDENTICAL	0	192under-exercised	torch.nn.functional.max_pool2d	cpu float32 cpu	MATCHED	-
linear_cpu	4x32->64	BIT_IDENTICAL	0	256under-exercised	torch.nn.functional.linear	cpu float32 cpu	MATCHED	-

Kernel

Shape

Runtime

max ULP

Floats

Oracle

Backend/Dev

Accuracy (vs truth)

Migration

Notes

sgemm_cpu

64x64

BIT_IDENTICAL

4,096

torch.matmul

cpu float32 cpu

MATCHED

sgemm_cpu

256x256

BIT_IDENTICAL

65,536

torch.matmul

cpu float32 cpu

MATCHED

sgemm_cpu

512x512

BIT_IDENTICAL

262,144

torch.matmul

cpu float32 cpu

MATCHED

relu_cpu

10000

BIT_IDENTICAL

10,000

torch.nn.functional.relu

cpu float32 cpu

MATCHED

silu_cpu

10000

BIT_IDENTICAL

10,000

torch.nn.functional.silu

cpu float32 cpu

MATCHED

mish_cpu

10000

BIT_IDENTICAL

10,000

torch.nn.functional.mish

cpu float32 cpu

MATCHED

fused_mm_bias_relu_cpu

256x256

BIT_IDENTICAL

65,536

torch.relu(x@W+b)

cpu float32 cpu

MATCHED

conv2d_cpu

1x3x16x16

BIT_IDENTICAL

2,048

torch.nn.functional.conv2d

cpu float32 cpu

MATCHED

upsample_cpu

1x8x4x4

BIT_IDENTICAL

512under-exercised

torch.nn.functional.interpolate

cpu float32 cpu

MATCHED

sigmoid_cpu

10000

BIT_IDENTICAL

10,000

torch.nn.functional.sigmoid

cpu float32 cpu

MATCHED

tanh_cpu

10000

BIT_IDENTICAL

10,000

torch.nn.functional.tanh

cpu float32 cpu

MATCHED

gelu_cpu

10000

PASS_ABS_TOLERANCE

127951

7,682 / 10,000

torch.nn.functional.gelu

cpu float32 cpu

IMPROVED
|err| 1.09e-08/3.86e-08 (3.5× better) · rounded 6,211/2,336

neg_cpu

10000

BIT_IDENTICAL

10,000

torch.neg

cpu float32 cpu

MATCHED

abs_cpu

10000

BIT_IDENTICAL

10,000

torch.abs

cpu float32 cpu

MATCHED

exp_cpu

10000

BIT_IDENTICAL

10,000

torch.exp

cpu float32 cpu

MATCHED

reduce_sum_cpu

1024

BIT_IDENTICAL

1under-exercised

torch.sum

cpu float32 cpu

MATCHED

reduce_mean_cpu

1024

BIT_IDENTICAL

1under-exercised

torch.mean

cpu float32 cpu

MATCHED

reduce_max_cpu

1024

BIT_IDENTICAL

1under-exercised

torch.max

cpu float32 cpu

MATCHED

softmax_cpu

32x64

BIT_IDENTICAL

2,048

torch.nn.functional.softmax

cpu float32 cpu

MATCHED

layernorm_cpu

8x128

BIT_IDENTICAL

1,024

torch.nn.functional.layer_norm

cpu float32 cpu

MATCHED

maxpool2d_cpu

1x3x16x16

BIT_IDENTICAL

192under-exercised

torch.nn.functional.max_pool2d

cpu float32 cpu

MATCHED

linear_cpu

4x32->64

BIT_IDENTICAL

256under-exercised

torch.nn.functional.linear

cpu float32 cpu

MATCHED

By Op — Two-Axis Join

Runtime axis × Migration axis, joined by canonical op per dashboard/OP_MAPPING.md. 7 both · 12 runtime-only · 9 emitted-only · 28 ops total. Different populations, overlap by operation.

Op	State	Runtime axis (vs oracle)	Migration axis (swipl→swilgt)
layer_norm	BOTH	layernorm_cpu BIT_IDENTICAL	k_layer_norm byte-identical
matmul	BOTH	sgemm_cpu BIT_IDENTICAL linear_cpu BIT_IDENTICAL	k_matmul byte-identical k_sgemv_substrate_native byte-identical k_sgemv_cublas_match byte-identical
relu	BOTH	relu_cpu BIT_IDENTICAL	k_relu byte-identical k_vecmat_relu byte-identical
sigmoid	BOTH	sigmoid_cpu BIT_IDENTICAL	k_sigmoid byte-identical
silu	BOTH	silu_cpu BIT_IDENTICAL	k_silu byte-identical k_vecmat_silu byte-identical
softmax	BOTH	softmax_cpu BIT_IDENTICAL	k_softmax byte-identical
tanh	BOTH	tanh_cpu BIT_IDENTICAL	k_tanh byte-identical
abs	RUNTIME ONLY	abs_cpu BIT_IDENTICAL	(no emitted counterpart)
conv2d	RUNTIME ONLY	conv2d_cpu BIT_IDENTICAL	(no emitted counterpart)
exp	RUNTIME ONLY	exp_cpu BIT_IDENTICAL	(no emitted counterpart)
fused_matmul_bias_relu	RUNTIME ONLY	fused_mm_bias_relu_cpu BIT_IDENTICAL	(no emitted counterpart)
gelu_erf	RUNTIME ONLY	gelu_cpu PASS_ABS_TOLERANCE (127951 ULP)	(no emitted counterpart)
max_pool2d	RUNTIME ONLY	maxpool2d_cpu BIT_IDENTICAL	(no emitted counterpart)
mish	RUNTIME ONLY	mish_cpu BIT_IDENTICAL	(no emitted counterpart)
neg	RUNTIME ONLY	neg_cpu BIT_IDENTICAL	(no emitted counterpart)
reduce_max	RUNTIME ONLY	reduce_max_cpu BIT_IDENTICAL	(no emitted counterpart)
reduce_mean	RUNTIME ONLY	reduce_mean_cpu BIT_IDENTICAL	(no emitted counterpart)
reduce_sum	RUNTIME ONLY	reduce_sum_cpu BIT_IDENTICAL	(no emitted counterpart)
upsample	RUNTIME ONLY	upsample_cpu BIT_IDENTICAL	(no emitted counterpart)
add	EMITTED ONLY	(no runtime cell)	k_add byte-identical
causal_mask	EMITTED ONLY	(no runtime cell)	k_causal_mask byte-identical
embed	EMITTED ONLY	(no runtime cell)	k_embed byte-identical
gelu_tanh	EMITTED ONLY	(no runtime cell)	k_gelu_tanh byte-identical k_vecmat_gelu byte-identical
mul	EMITTED ONLY	(no runtime cell)	k_mul byte-identical
rms_norm	EMITTED ONLY	(no runtime cell)	k_rms_norm byte-identical
rope	EMITTED ONLY	(no runtime cell)	k_rope byte-identical
scale	EMITTED ONLY	(no runtime cell)	k_scale byte-identical
vecmat	EMITTED ONLY	(no runtime cell)	k_vecmat byte-identical

State

Runtime axis (vs oracle)

Migration axis (swipl→swilgt)

layer_norm

BOTH

layernorm_cpu BIT_IDENTICAL

k_layer_norm byte-identical

matmul

BOTH

sgemm_cpu BIT_IDENTICAL

linear_cpu BIT_IDENTICAL

k_matmul byte-identical

k_sgemv_substrate_native byte-identical

k_sgemv_cublas_match byte-identical

relu

BOTH

relu_cpu BIT_IDENTICAL

k_relu byte-identical

k_vecmat_relu byte-identical

sigmoid

BOTH

sigmoid_cpu BIT_IDENTICAL

k_sigmoid byte-identical

silu

BOTH

silu_cpu BIT_IDENTICAL

k_silu byte-identical

k_vecmat_silu byte-identical

softmax

BOTH

softmax_cpu BIT_IDENTICAL

k_softmax byte-identical

tanh

BOTH

tanh_cpu BIT_IDENTICAL

k_tanh byte-identical

abs

RUNTIME ONLY

abs_cpu BIT_IDENTICAL

(no emitted counterpart)

conv2d

RUNTIME ONLY

conv2d_cpu BIT_IDENTICAL

(no emitted counterpart)

exp

RUNTIME ONLY

exp_cpu BIT_IDENTICAL

(no emitted counterpart)

fused_matmul_bias_relu

RUNTIME ONLY

fused_mm_bias_relu_cpu BIT_IDENTICAL

(no emitted counterpart)

gelu_erf

RUNTIME ONLY

gelu_cpu PASS_ABS_TOLERANCE (127951 ULP)

(no emitted counterpart)

max_pool2d

RUNTIME ONLY

maxpool2d_cpu BIT_IDENTICAL

(no emitted counterpart)

mish

RUNTIME ONLY

mish_cpu BIT_IDENTICAL

(no emitted counterpart)

neg

RUNTIME ONLY

neg_cpu BIT_IDENTICAL

(no emitted counterpart)

reduce_max

RUNTIME ONLY

reduce_max_cpu BIT_IDENTICAL

(no emitted counterpart)

reduce_mean

RUNTIME ONLY

reduce_mean_cpu BIT_IDENTICAL

(no emitted counterpart)

reduce_sum

RUNTIME ONLY

reduce_sum_cpu BIT_IDENTICAL

(no emitted counterpart)

upsample

RUNTIME ONLY

upsample_cpu BIT_IDENTICAL

(no emitted counterpart)

add

EMITTED ONLY

(no runtime cell)

k_add byte-identical

causal_mask

EMITTED ONLY

(no runtime cell)

k_causal_mask byte-identical

embed

EMITTED ONLY

(no runtime cell)

k_embed byte-identical

gelu_tanh

EMITTED ONLY

(no runtime cell)

k_gelu_tanh byte-identical

k_vecmat_gelu byte-identical

mul

EMITTED ONLY

(no runtime cell)

k_mul byte-identical

rms_norm

EMITTED ONLY

(no runtime cell)

k_rms_norm byte-identical

rope

EMITTED ONLY

(no runtime cell)

k_rope byte-identical

scale

EMITTED ONLY

(no runtime cell)

k_scale byte-identical

vecmat

EMITTED ONLY

(no runtime cell)

k_vecmat byte-identical

1 open cell(s) toward full bit-perfect · LlamaTov / bpd-substrate · schema: dashboard/CONGRUENCE_SCHEMA.md