[danielhanchen]

Adds support for Reflection's Beam-1.0-501B-A23B (HF BeamForCausalLM): converter, model graph and a chat parser for its tool-call format. Opening as a draft because the weights are not public yet, so the real-model results below cannot be reproduced from outside at the moment.

Architecture

subsystem	how it is handled
MoE, 256 experts, top-8, sigmoid routing with grouped selection	build_moe_ffn; correction bias used for selection only, routed scale and normalised weights required in the GGUF
3 leading dense SwiGLU layers	standard FFN path
Interleaved sliding-window (512) and NoPE full-attention layers	iSWA cache; RoPE (NeoX) on the sliding layers only
Power-law query scale on the NoPE layers	optional form of llm_graph_input_attn_temp, off by default
q/k RMSNorm, per-head sigmoid output gate, scaled residuals, final logit softcap	in src/models/beam.cpp
Tool calls (<|channel|> + <|constrain|>xml + rfl:parameter)	new harmony-xml chat parser, detected only when all three markers are present

No new ggml op.

Correctness

• Reference parity on a tiny random checkpoint with the real layer pattern, tokenizer and routing constants, against the transformers implementation in fp32: top-1 agreement and selected-expert Jaccard 1.0 on 64, 600 and 2000 token prompts; max logit difference 6.8e-5 / 5.2e-4 / 2.0e-3.
• test-llama-archs -a beam: OK on CPU, CUDA and Meta. test-chat, test-jinja pass.
• Real weights: the 501B checkpoint converts, and quantized GGUFs load and generate on CUDA. Chat template renders byte-identical to apply_chat_template; multi-turn tool-call sessions (parallel calls, typed and nested arguments, long logs) pass end to end through llama-server.
• Malformed GGUFs (missing routing keys, no sliding layer, zero temperature scale) fail at load instead of at the first decode.

Effect on other architectures

• llama-graph.cpp: the power-law temperature is a separate branch; existing users take the old path. llm_graph_input_attn_temp gains a shape-only can_reuse, since set_input recomputes the scale every ubatch.
• llama-vocab.cpp: Beam's </s> stays a normal token (it is plain text in its tokenizer), using the existing gemma4/plamo exception keyed on a Beam-only token. Token ids, EOG sets and attributes are unchanged on 19 upstream test vocabs and 224 local GGUFs.
• common/jinja: the list filter accepts objects (returns the keys).
• common/chat.cpp: the new parser is only selected for templates containing all three markers.

AI Usage

Used Claude and Local Models for testing, iteration and code design - manual verification of model / PR usage

[danielhanchen]

GGUFs at https://huggingface.co/unsloth/Beam-1.0-501B-A23B-GGUF for those wanting to try them!

[pwilkin]

The CUDA error is transient, will merge after all the remaining checks are OK.