Analysis outputs

analyze(doc, options) returns every number at once. validate(doc, options) returns the same thing with design-rule findings attached.

Nothing in the UI computes its own numbers: one call per document and operating point feeds every panel.

The operating point

Conditions the design is measured under, never properties of it:

Option	Meaning
B	Micro-batch size
T	Sequence length in tokens
bytes	Bytes per element of the activation dtype
hardware	GPU profile id
gpus	How many
sharding	ZeRO stage, tensor and pipeline parallel degree
recompute	none / selective / full
precision	How a bf16 or fp16 dtype is trained. mixed, the default: bf16 weights and activations over an fp32 master copy, as Megatron does it. autocast: torch.autocast, fp32 weights cast to bf16 at each matrix multiply. It keeps the residual stream and the norms in fp32 and a bf16 copy of every weight for the backward pass (reported as memory.train.castWeights), and splits the same 16 bytes a parameter 4/4/8 rather than 2/2/12. Ignored in fp32
tokens	Training token budget, for cost and Chinchilla
packing	Training rows packed with documents, { mean, spread }: the mean length in tokens and the coefficient of variation, 0 for fixed lengths and 1 for exponential. Absent is one document a row

params

total, active, byPath, byCategory.

Sparse designs differ in the two totals: a container carries a total multiplier that drives the parameter count and an active multiplier that drives FLOPs. That single mechanism is what makes mixture-of-experts work.

byPath holds leaf blocks only. A container's contribution is the sum of the paths beneath it.

flops

Per token, unless stated.

Field	Meaning
fwdDense	Matmul FLOPs that do not depend on sequence length
fwdAttention	Score and value products at the given T
fwdAttentionUnmasked	The same counted as if nothing were masked
fwdTotal	fwdDense + fwdAttention — what a fused causal kernel does
fwdTotalUnmasked	What a profiler reports
elementwise	Norms, activations, RoPE, residual adds — memory-bound, excluded above
trainPerToken	Forward plus backward, plus recomputation
attentionShare	Fraction of forward FLOPs inside attention
ruleOfThumb2N / ruleOfThumb6N	The usual approximations, for comparison
perStream	A design with two sequences only: each one's forward pass per token of its own, S then T
fwdPerExample	A design with two sequences only: one example's forward pass, both sequences' tokens
packed	Training under the operating point's packing, for a design whose mask keeps documents apart: fwdAttention, fwdTotal, trainPerToken and attentionShare, and fwdAttentionBlocks, the attention as a block-sparse kernel computes it, every 128 × 128 block holding a kept score computed whole

With a packing, every figure above except packed is still one document a row, which is what serving is. packed is training, and the training time and cost are counted from it. It is present only when a mask reads the documents, because a packing changes nothing else.

For a design with a second sequence every per-token figure above is per target token: the source's blocks are measured at S, and their share is spread over the target's tokens. options.S then says what S was.

The two totals exist because they are answers to different questions. See Verify a design against PyTorch.

For a convolutional design the token is one image, so T is 1 and every per-token figure reads as per-image.

kv

bytesPerToken, and the total at the current batch and sequence length. A state-space layer keeps a fixed state per sequence rather than one growing with every token, which is most of why hybrid stacks are cheap to serve.

memory

Weights, gradients, optimiser state and activations, before and after sharding, per GPU and in total. Activation figures follow the Megatron formulas.

Activation memory is attributed to tensors, not blocks: a block declares which inputs it must keep alive, and each producing tensor is counted once even when several blocks read it.

throughput

A roofline estimate: arithmetic intensity against the chosen hardware's compute and bandwidth, and which of the two binds.

cost and chinchilla

Training cost at the given token budget, and how far that budget is from compute-optimal — about twenty tokens per parameter.

symbols, infer, expanded, flat

The intermediate results, exposed because the UI needs them: the resolved symbol table, shapes before and after composite expansion, and the flattened list of primitive instances the totals are summed from.