From a79ab8f34222e1e0142a30eaa97e78ad077abca9 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Sat, 1 Mar 2025 08:25:27 +0200 Subject: Reduce size of compute buffers (#237) * This reduces compute buffer size for MLA * This should accomplish it for standard attention * Much better * Better concat for contiguous tensors If all the op does is to concatenate the second tensor to the first, why would we want to have a loop? --------- Co-authored-by: Iwan Kawrakow --- common/common.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) (limited to 'common/common.h') diff --git a/common/common.h b/common/common.h index ef5175f3..f35f3558 100644 --- a/common/common.h +++ b/common/common.h @@ -175,7 +175,8 @@ struct gpt_params { bool simple_io = false; // improves compatibility with subprocesses and limited consoles bool cont_batching = true; // insert new sequences for decoding on-the-fly bool flash_attn = false; // flash attention - int mla_attn = false; // MLA 0: standard attention, 1: MLA with K and transposed V cache, 2: MLA with just K cache + int mla_attn = 0; // MLA 0: standard attention, 1: MLA with K and transposed V cache, 2: MLA with just K cache + int attn_max_batch = 0; // Max batch size to use when computing attention (only applicable if flash_attn = false) bool fused_moe_up_gate = false; // fused up*unary(gate) op for MoE models bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix -- cgit v1.2.3