ggml : deprecate GGML_TASK_FINALIZE

[ggerganov]

The initial design of the compute tasks in ggml was each one to have 3 separate stages:

• GGML_TASK_INIT
• GGML_TASK_COMPUTE
• GGML_TASK_FINALIZE

So far, the GGML_TASK_FINALIZE step has been left completely unused and it seems it won't find any applications in the future. Therefore, it is best to remove it all together. This will simplify the code a little bit and also bring performance improvements as there will be fewer thread synchronization points during the graph computation

Related:

• Numa llama.cpp#1556

[rlanday]

Hi, I took a look at this and found that GGML_TASK_FINALIZE is used by ggml_compute_forward_cross_entropy_loss_f32():

https://github.com/ggerganov/llama.cpp/blob/178b1850ebd21b349cebbee887950e435c5aa2d3/ggml.c#L15453

Please advise.

[ggerganov]

INIT and FINALIZE are now optional and disabled for the majority of ops:

ggml/src/ggml.c

Lines 4080 to 4116 in 79ea7a4

	// WARN:
	// Mis-confguration can lead to problem that's hard to reason about:
	// * At best it crash or talks nosense.
	// * At worst it talks slightly difference but hard to perceive.
	//
	// An op has to enable INIT or FINALIZE when any of it's branch needs that pass.
	// Take care about compile options (e.g., GGML_USE_xxx).
	static bool GGML_OP_HAS_INIT [GGML_OP_COUNT] = { 0 };
	static bool GGML_OP_HAS_FINALIZE[GGML_OP_COUNT] = { 0 };
	static void ggml_setup_op_has_task_pass(void) {
	{ // INIT
	bool * p = GGML_OP_HAS_INIT;
	p[GGML_OP_ACC ] = true;
	p[GGML_OP_MUL_MAT ] = true;
	p[GGML_OP_OUT_PROD ] = true;
	p[GGML_OP_SET ] = true;
	p[GGML_OP_GET_ROWS_BACK ] = true;
	p[GGML_OP_DIAG_MASK_INF ] = true;
	p[GGML_OP_DIAG_MASK_ZERO ] = true;
	p[GGML_OP_CONV_1D ] = true;
	p[GGML_OP_CONV_2D ] = true;
	p[GGML_OP_CONV_TRANSPOSE_2D ] = true;
	p[GGML_OP_FLASH_ATTN_BACK ] = true;
	p[GGML_OP_CROSS_ENTROPY_LOSS ] = true;
	p[GGML_OP_ADD_REL_POS ] = true;
	}
	{ // FINALIZE
	bool * p = GGML_OP_HAS_FINALIZE;
	p[GGML_OP_CROSS_ENTROPY_LOSS ] = true;
	}
	}

This is good enough for now