LFM2.5-8B-A1B-FP8 (block-FP8, deep_gemm)

Block-FP8 (e4m3, weight_block_size=[128,128], dynamic activations) quant of LiquidAI/LFM2.5-8B-A1B, byte-faithful to the Qwen/Qwen3.6-35B-A3B-FP8 recipe. Serve with SGLang --quantization fp8 → deep_gemm on H100/SM90.

NOT compressed-tensors — deliberately avoids the per-channel path that made the prior cyankiwi/LFM2.5-8B-A1B-AWQ-FP8 checkpoint emit degenerate whitespace.

Skipped (kept bf16): embeddings, lm_head, all RMSNorms, ShortConv (conv.in_proj/out_proj/conv1d), MoE router gate, and any layer not 128-divisible.

Calibration: domain-matched (used for the coherence/validation gate; the weight cast itself is data-free per-block absmax).

Cast stats: {"quantized_linears": 2142, "skipped_modules": 60, "shards": 3, "quant_GB": 7.9, "skip_GB": 1.13, "total_GB": 9.04, "skip_sample": ["lm_head", "model.embed_tokens", "model.layers.0.conv.in_proj", "model.layers.0.conv.out_proj", "model.layers.1.conv.in_proj", "model.layers.1.conv.out_proj", "model.layers.10.feed_forward.gate", "model.layers.11.conv.in_proj", "model.layers.11.conv.out_proj", "model.layers.11.feed_forward.gate", "model.layers.12.conv.in_proj", "model.layers.12.conv.out_proj"]}