vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : add Qwen2VL support + multimodal RoPE (#10361)

Browse source 
* Barebone Qwen2VL LLM convertor

* Add Qwen2VL cli entrypoint

* [WIP] add qwen2vl arch

* Verify m-rope output

* Add vl-rope/2d-rope support for qwen2vl ViT

* update qwen2vl cli tool

* update 5D tensor op workaround

* [WIP] qwen2vl vision model

* make batch and clip utils compatible with qwen2vl

* [WIP] create inference workflow, gguf convert script but fix

* correcting vision-rope behavior, add the missing last layer back to ViT

* add arg parser to qwen2vl_surgery

* replace variable size array with vector

* cuda-gdb cmake preset

* add fp32 mrope, vision rope kernel

* add fp16 support for qwen2vl and m-rope

* add `GGML_ROPE_TYPE_MROPE`, `GGML_ROPE_TYPE_VISION`

* fix rope op mode switching, out dated func args

* update `llama_hparams`

* update to keep up stream changes

* resolve linter, test errors

* add makefile entry, update speical image padding token

* add mrope unit test, fix few compiler warnings

* rename `mrope` related function, params

* minor updates on debug util, bug fixs

* add `m-rope` testcase to `test-backend-ops`

* Apply suggestions from code review

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* fix traililng whitespce

* store `llama_hparams.rope_sections` with fixed size array

* update position id tensor size check in GGML_OP_ROPE

* minor updates

* update `ggml_backend_*_supports_op` of unsupported backends

* remote old `rope_section` compare operator

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
HimariO 2024-12-14 20:43:46 +08:00 • committed by GitHub
parent 56eea0781c
commit ba1cb19cdd
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
24 changed files with 1873 additions and 114 deletions
Download patch file Download diff file Expand all files Collapse all files

52
ggml/src/ggml.c
View file

@ -3517,15 +3517,18 @@ static struct ggml_tensor * ggml_rope_impl(
GGML_ASSERT(c->ne[0] >= n_dims / 2);
}
int sections[4] = {0, 0, 0, 0};
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
memcpy(params + 5, &freq_base, sizeof(float));
memcpy(params + 6, &freq_scale, sizeof(float));
memcpy(params + 7, &ext_factor, sizeof(float));
memcpy(params + 8, &attn_factor, sizeof(float));
memcpy(params + 9, &beta_fast, sizeof(float));
memcpy(params + 10, &beta_slow, sizeof(float));
memcpy(params + 11, &sections, sizeof(int)*4);
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_ROPE;
@ -3547,6 +3550,53 @@ struct ggml_tensor * ggml_rope(
);
}
struct ggml_tensor * ggml_rope_multi(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
int n_dims,
int sections[4],
int mode,
int n_ctx_orig,
float freq_base,
float freq_scale,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow) {
// Multimodal Rotary Position Embedding
GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported");
GGML_ASSERT(ggml_is_vector(b));
GGML_ASSERT(b->type == GGML_TYPE_I32);
GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
if (c) {
GGML_ASSERT(c->type == GGML_TYPE_F32);
GGML_ASSERT(c->ne[0] >= n_dims / 2);
}
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
int32_t params[11 + 4] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
memcpy(params + 5, &freq_base, sizeof(float));
memcpy(params + 6, &freq_scale, sizeof(float));
memcpy(params + 7, &ext_factor, sizeof(float));
memcpy(params + 8, &attn_factor, sizeof(float));
memcpy(params + 9, &beta_fast, sizeof(float));
memcpy(params + 10, &beta_slow, sizeof(float));
memcpy(&params[11], sections, sizeof(int)*4);
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_ROPE;
result->src[0] = a;
result->src[1] = b;
result->src[2] = c;
return result;
}
struct ggml_tensor * ggml_rope_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,