vllm.models.mistral.mistral_large4.amd.specialized_decode.mono.geometry
¶
Shared compile-time layouts and schedules for fused model-layer kernels.
Functions:
-
dn_tile–Return hidden rows per expert-down / FFN peer-reduce task.
-
layout–Return scratch and double-buffered symmetric-region byte offsets.
-
stage_tasks–Return
(stage name, task count)pairs in execution order.
dn_tile(samples, hidden=ML4_CONFIG.hidden)
¶
Return hidden rows per expert-down / FFN peer-reduce task.
layout(samples, heads, npes, sparse_attention_topk, moe_mode=MoeMode.W8A8, model_config=ML4_CONFIG, attention_only=False, dedicated_input_norm=False)
¶
Return scratch and double-buffered symmetric-region byte offsets.
Source code in vllm/models/mistral/mistral_large4/amd/specialized_decode/mono/geometry.py
stage_tasks(samples, heads, sparse_attention_topk, model_config=ML4_CONFIG, attention_only=False, dedicated_input_norm=False)
¶
Return (stage name, task count) pairs in execution order.