FP8 Scale Denominator Clamp | Generated by AI

Home 2026.07

Question: In nanochat’s FP8 quantization code, what does .clamp(min=EPS) do in the line scale = fp8_max / amax.double().clamp(min=EPS)?

Answer:

This line is from nanochat’s hand-rolled FP8 quantization (replacing the torchao dependency — the commit deletes the torchao dependency and creates a custom API-matched version). Full context:

scale = fp8_max / amax.double().clamp(min=EPS)
scale = scale.float()
x_scaled = x.float() * scale
x_clamped = x_scaled.clamp(-fp8_max, fp8_max)
x_fp8 = x_clamped.to(fp8_dtype)
inv_scale = scale.reciprocal()

amax = the tensor’s absolute max (x.abs().max()), computed per-tensor or per-row depending on the recipe (tensorwise vs rowwise). It tells you the dynamic range you need to fit into FP8’s tiny range (e.g. e4m3 has fp8_max ≈ 448).

.clamp(min=EPS) here is a divide-by-zero guard, not a saturation clamp. Two different clamps show up in this snippet and they do opposite jobs:

  1. amax.double().clamp(min=EPS) — floors the denominator. If a tensor/row is all zeros (dead activations, a masked-out row, a zeroed gradient), amax = 0, and fp8_max / 0 → inf/nan. Clamping the denominator to a tiny epsilon (EPS, typically something like 1e-12) forces scale to just be a large finite number instead of inf. Downstream, x_scaled = 0 * huge_scale = 0, so the zero tensor correctly quantizes to zero rather than propagating NaNs through the matmul.

  2. x_scaled.clamp(-fp8_max, fp8_max) — this is the actual saturating clamp on the numerator, done for the opposite reason: .to(fp8_dtype) in PyTorch wraps/overflows rather than saturating when a float exceeds the target dtype’s representable range (comment literally says this — clamp prevents overflow when casting since PyTorch’s default is to wrap, not saturate). Without this second clamp, an outlier value could wrap around to a huge value with the wrong sign, silently corrupting the matmul.

So: the first .clamp(min=EPS) protects the scale computation from a degenerate all-zero input; the second .clamp(-fp8_max, fp8_max) protects the cast from silent wraparound on outliers. Both are classic “guard the numerics at the boundary” moves you’ll see in any quantization codebase (torchao, transformer_engine, bitsandbytes all have equivalent epsilon floors on the amax denominator).

One implementation detail worth noting: amax.double() — doing the reciprocal-scale math in fp64 before casting scale back to fp32 avoids precision loss when amax is very small, since fp8_max / amax can itself be numerically sensitive near the epsilon floor.

References:

Back Donate