mstar.model.glm52.config

GLM-5.2 architecture + generation config.

Classes

Glm52ModelConfig([vocab_size, hidden_size, ...])

class mstar.model.glm52.config.Glm52ModelConfig(vocab_size: int = 154880, hidden_size: int = 6144, num_hidden_layers: int = 78, num_attention_heads: int = 64, rms_norm_eps: float = 1e-05, hidden_act: str = 'silu', tie_word_embeddings: bool = False, q_lora_rank: int = 2048, kv_lora_rank: int = 512, qk_nope_head_dim: int = 192, qk_rope_head_dim: int = 64, v_head_dim: int = 256, mla_absorb: bool = True, index_n_heads: int = 32, index_head_dim: int = 128, index_topk: int = 2048, index_topk_freq: int = 4, index_skip_topk_offset: int = 3, index_share_for_mtp_iteration: bool = True, indexer_rope_interleave: bool = True, dsa_long_context: bool = False, first_k_dense_replace: int = 3, intermediate_size: int = 12288, moe_intermediate_size: int = 2048, n_routed_experts: int = 256, n_shared_experts: int = 1, num_experts_per_tok: int = 8, routed_scaling_factor: float = 2.5, scoring_func: str = 'sigmoid', topk_method: str = 'noaux_tc', norm_topk_prob: bool = True, moe_router_dtype: str = 'float32', rope_theta: float = 8000000.0, rope_interleave: bool = True, max_position_embeddings: int = 1048576, max_seq_len: int = 2048, num_nextn_predict_layers: int = 1, eos_token_ids: tuple[int, ...] = (154820, 154827, 154829), pad_token_id: int = 154820, max_output_tokens: int | None = None, temperature: float = 1.0, top_p: float = 0.95, repetition_penalty: float = 1.0, ignore_eos: bool = False, quantization_config: mstar.utils.fp8_block.Fp8BlockQuantConfig | None = None, moe_fp8_resident: bool = True, moe_quant_kernel: str = 'auto', graph_compile: bool = True, prefill_token_buckets: list[int] | None = None, prefill_capture_batch_sizes: list[int] | None = None)

Bases: object

Parameters:

• vocab_size (int)

• hidden_size (int)

• num_hidden_layers (int)

• num_attention_heads (int)

• rms_norm_eps (float)

• hidden_act (str)

• tie_word_embeddings (bool)

• q_lora_rank (int)

• kv_lora_rank (int)

• qk_nope_head_dim (int)

• qk_rope_head_dim (int)

• v_head_dim (int)

• mla_absorb (bool)

• index_n_heads (int)

• index_head_dim (int)

• index_topk (int)

• index_topk_freq (int)

• index_skip_topk_offset (int)

• index_share_for_mtp_iteration (bool)

• indexer_rope_interleave (bool)

• dsa_long_context (bool)

• first_k_dense_replace (int)

• intermediate_size (int)

• moe_intermediate_size (int)

• n_routed_experts (int)

• n_shared_experts (int)

• num_experts_per_tok (int)

• routed_scaling_factor (float)

• scoring_func (str)

• topk_method (str)

• norm_topk_prob (bool)

• moe_router_dtype (str)

• rope_theta (float)

• rope_interleave (bool)

• max_position_embeddings (int)

• max_seq_len (int)

• num_nextn_predict_layers (int)

• eos_token_ids (tuple[int, ...])

• pad_token_id (int)

• max_output_tokens (int | None)

• temperature (float)

• top_p (float)

• repetition_penalty (float)

• ignore_eos (bool)

• quantization_config (Fp8BlockQuantConfig | None)

• moe_fp8_resident (bool)

• moe_quant_kernel (str)

• graph_compile (bool)

• prefill_token_buckets (list[int] | None)

• prefill_capture_batch_sizes (list[int] | None)

cache_latent_dim: int

dsa_long_context: bool = False

eos_token_ids: tuple[int, ...] = (154820, 154827, 154829)

first_k_dense_replace: int = 3

classmethod from_hf_config(hf_config)

Geometry from the checkpoint’s config.json; serving knobs keep their defaults.

Parameters:

hf_config (dict)

Return type:

Glm52ModelConfig

graph_compile: bool = True

hidden_act: str = 'silu'

hidden_size: int = 6144

ignore_eos: bool = False

index_head_dim: int = 128

index_n_heads: int = 32

index_share_for_mtp_iteration: bool = True

index_skip_topk_offset: int = 3

index_topk: int = 2048

index_topk_freq: int = 4

indexer_rope_interleave: bool = True

intermediate_size: int = 12288

kv_lora_rank: int = 512

max_output_tokens: int | None = None

max_position_embeddings: int = 1048576

property max_prompt_tokens: int

The longest prompt served. Decode runs at least one step after the prefill and the next may already be scheduled when it stops, so the prompt leaves two rows under the context limit; a DSA prefill attends densely, within index_topk.

max_seq_len: int = 2048

mla_absorb: bool = True

moe_fp8_resident: bool = True

moe_intermediate_size: int = 2048

moe_quant_kernel: str = 'auto'

moe_router_dtype: str = 'float32'

n_routed_experts: int = 256

n_shared_experts: int = 1

norm_topk_prob: bool = True

num_attention_heads: int = 64

property num_dense_layers: int

num_experts_per_tok: int = 8

num_hidden_layers: int = 78

num_nextn_predict_layers: int = 1

pad_token_id: int = 154820

property padded_head_dim: int

Naive-MLA q/k/v pad target; FlashInfer paged kernels require 64/128/256.

position_config()

The positions over the KV cache and the RoPE the layers rotate with: interleaved over qk_rope_head_dim, from an fp32 cos/sin table built once (the angles as the reference computes them, for every position up to max_seq_len).

Return type:

PositionConfig

prefill_capture_batch_sizes: list[int] | None = None

prefill_token_buckets: list[int] | None = None

q_lora_rank: int = 2048

property qk_head_dim: int

qk_nope_head_dim: int = 192

qk_rope_head_dim: int = 64

quantization_config: Fp8BlockQuantConfig | None = None

classmethod reduced()

Tiny-dim variant for CPU tests: same shapes family, random weights.

Return type:

Glm52ModelConfig

classmethod reduced_fp8(block=(16, 16))

Parameters:

block (tuple[int, int])

Return type:

Glm52ModelConfig

repetition_penalty: float = 1.0

rms_norm_eps: float = 1e-05

rope_interleave: bool = True

rope_theta: float = 8000000.0

routed_scaling_factor: float = 2.5

scoring_func: str = 'sigmoid'

property softmax_scale: float

MLA’s attention scale. Plain RoPE, so no YaRN mscale folds in.

temperature: float = 1.0

tie_word_embeddings: bool = False

top_p: float = 0.95

topk_method: str = 'noaux_tc'

v_head_dim: int = 256

vocab_size: int = 154880