v0.3.2a2

Adds narrower recomputation, two lookahead options and web UI fields for every long-context switch, and fixes a crash found by training on the real corpus. Every new switch is off by default. reports/long_context_20261109/real_corpus_ab.json holds the measurements below.

Added

• recompute_scope="mlp": with activation_checkpointing, recompute only the MoE/FFN of each block ("block" recomputes the whole block, as before). At 32k on the real corpus it saves 1.4 GB (5%) and loses 11% speed, against 6.9 GB (25%) and 19% for whole blocks, so it suits runs that need only a little more memory.
• mla_upproj_recompute: keep only the MLA latent and rebuild K and V per attention block. Exact, but it saved no memory (24.47 -> 24.29 GiB at 64k) and costs 6.7% speed, so it stays off.
• lookahead_input_detach (the lookahead router reads a gradient-free hidden state) and temporal_grad_span (cut the temporal-locality gradient chain every n positions). Both are exact in the forward pass. On the en-zh data they were neutral (validation CE within 1e-3), so they stay off.
• Web UI: every long-context field (attention_block_size, ce_chunk_size, moe_chunk_tokens, recompute_scope, recompute_policy, mla_upproj_recompute, document_masking, reset_router_at_documents, reset_lookahead_at_documents, lookahead_input_detach, temporal_grad_span) with labels and help in all four languages.

Fixed

• reset_router_at_documents and document_masking raised "Attempting to eval an array without a primitive" inside the compiled LongRun step, because the document ids were built with NumPy from a traced array. chronos.expert_clusters.segment_ids_mx builds them in MLX, and a regression test runs the option through mx.compile.
• reset_lookahead_at_documents="false" (the string the web UI sends) was read as true. Strings are now parsed; the empty string follows the router switch.

Measured on the real en-zh corpus (production 50M config, random init, M5 Max, bf16)

long = attention_block_size=1024, ce_chunk_size=1024, moe_chunk_tokens=8192. Validation CE differs by at most 3e-3 between arms.

Context	Default: tokens/s, peak	long: tokens/s, peak
8k	8.9k, 40.6 GiB	19.5k, 7.5 GiB
16k	1.9k, 141.5 GiB (over this machine's 128 GB)	15.2k, 13.8 GiB
32k	not run	8.8k, 27.0 GiB
64k	not run	3.7k, 54.6 GiB

Larger attention blocks speed up one MLA layer (0.62 s -> 0.53 s at 64k with blocks of 8192) but slow the whole model, because the sliding-window layers get slower (9.9 s -> 16.6 s per step); 1024 stays the recommendation.