Training speculative decoding draft heads at scale introduces two bottlenecks: the memory cost of the language-modeling loss and the attention pattern used by EAGLE-style draft heads. Streaming Cross Entropy avoids materializing full vocabulary logits for every token at once, while block-sparse FlashAttention lets EAGLE-3 draft-head training use efficient attention kernels. This engineering layer supports the Custom Speculator Training direction in Nebius Token Factory, helping teams train workload-specific drafters and evaluate them against production traffic. Unlike generic long-context fine-tuning, draft-head training adds extra token-prediction objectives, where each head may predict a different future token, making target alignment, masking, loss computation, and attention efficiency more delicate. The logits bottleneck arises because the final language-modeling loss path—hidden states to LM head to logits to Cross Entropy—can materialize the full logits tensor. For Llama 3.1 8B with batch_size = 16, sequence_length = 32k, vocab_size = 128256, and logits dtype = fp32, the logits tensor alone requires 16 × 32768 × 128256 × 4 bytes ≈ 250.5 GiB, while hidden states require only 16 × 32768 × 4096 × 2 bytes ≈ 4 GiB. If both logits and gradient are materialized, peak memory roughly doubles to ~501GiB. This issue affects both small and large models