fix(tools): clamp block max in block_fp8 to avoid NaN weights from all-zero blocks - #2256

Problem

block_fp8 — the default quantization strategy of tools/convert_hf_to_fp8.py — computes

scale = block_max.to(torch.float32) / FP8_MAX
qweight = (qweight / scale).clamp(...)

without clamping the block max away from zero, unlike channel_fp8 and tensor_fp8 in the same file, which both do .clamp(min=1e-12).

For an all-zero 128×128 block (padding vocab rows, unused MoE experts, zeroed gate weights — all present in large MoE checkpoints), block_max == 0 gives scale == 0 and qweight == 0 / 0 == NaN (torch.clamp(NaN) stays NaN). The NaN weights are silently written into the converted safetensors, and dequantization (NaN * 0 = NaN) then poisons the forward pass of the converted model. Near-zero blocks similarly yield inf, which gets clamped to ±448 — a wrong value with no error raised.

Reproduction (CPU):

w = torch.zeros(256, 256, dtype=torch.bfloat16)
w[0, 0] = 1.0  # only the first 128x128 tile is non-zero
q, s = block_fp8(w, (128, 128))
torch.isnan(q.float()).sum()  # 49152 NaNs on current main

Fix

Clamp the block max to 1e-12, exactly matching channel_fp8 / tensor_fp8:

scale = block_max.clamp(min=1e-12).to(torch.float32) / FP8_MAX

An all-zero block then quantizes to zeros with a tiny scale, and dequantizes back to exact zeros. Non-zero blocks are completely unaffected.

Tests

Added tests/test_block_fp8_zero_block.py (CPU, @pytest.mark.unit):

• mixed zero/non-zero tiles → no NaN/inf in the output, all scales > 0 (fails without the fix);
• all-zero block round-trips to exact zeros (fails without the fix);
• non-zero blocks quantize/dequantize within FP8 tolerance (passes before and after — guards against behavior change).

All 3 tests pass locally, plus ruff and black --check are clean.