bitnet-b1.58-2B-4T runs FFN with SiLU instead of relu2 - wrong logits on every backend

从 src/models/bitnet.cpp 和 src/llama-model.cpp 开始，然后检查 src/llama-graph.cpp 以及两个 BitNet 转换脚本中对隐藏激活的处理。使用 gguf-dump 验证转换后的 GGUF，并比较模型 microsoft/bitnet-b1.58-2B-4T 的 logits 或生成文本。完成的标志是记录并使用 relu2，同时较旧的 GGUF 文件保留 SiLU 行为。

bitnet-b1.58-2B-4T runs the FFN with SiLU instead of relu² — wrong logits on every backend

Summary

The official microsoft/bitnet-b1.58-2B-4T model declares "hidden_act": "relu2" in its config.json, but the BitNet.cpp runtime never reads it: the FFN graph is built with a hardcoded LLM_FFN_SILU. The model therefore runs with the wrong activation function on every backend (CPU and GPU), producing wrong-but-finite output — the same failure class as #586.

Root cause

• src/models/bitnet.cpp:133 — build_ffn(...) is called with a hardcoded LLM_FFN_SILU.
• src/llama-model.cpp — the LLM_FFN_OP_TYPES_FROM_STRING map has "relu" but no "relu2", and llm_ffn_op_type_from_string() is never wired into the bitnet arch.
• utils/convert-hf-to-gguf-bitnet.py / utils/convert-ms-to-gguf-bitnet.py — neither converter writes <arch>.hidden_activation, so the existing plumbing (llama_hparams::llm_ffn_op, LLM_KV_HIDDEN_ACT, LLM_FFN_RELU_SQR in src/llama-graph.cpp) is never exercised for bitnet models.

The graph builder already implements LLM_FFN_RELU_SQR (relu → sqr, src/llama-graph.cpp:1690); with LLM_FFN_PAR gating it produces exactly relu²(gate) * up, matching the HF BitnetMLP (act_fn(gate) * up).

Fix (attached patches)

1. src/llama-model.cpp: add { "relu2", LLM_FFN_RELU_SQR } to LLM_FFN_OP_TYPES_FROM_STRING.
2. src/models/bitnet.cpp: read <arch>.hidden_activation in load_arch_hparams (mirroring llama_model_modern_bert), defaulting to LLM_FFN_SILU for backward compatibility with existing GGUF files; use hparams.llm_ffn_op in the graph.
3. Converters: write add_hidden_act("relu2").

Verification

• Convert microsoft/bitnet-b1.58-2B-4T with the patched converter; gguf-dump should show bitnet_b158.hidden_activation = "relu2".
• Compare logits before/after: with the fix the model produces coherent text; without it, output is wrong-but-finite.
• Old GGUF files (no key) keep the previous SiLU behavior — no regression.

Related

• #586 — same wrong-but-finite class (ARM i2_s layout).