Skip to content

Commit

Permalink
Add initial SD3 lcpp patch + instruction
Browse files Browse the repository at this point in the history 
#136
  • Loading branch information
@city96
city96 committed Oct 22, 2024
1 parent 0ef158d commit 4d7bb93
Show file tree
Hide file tree
Showing 2 changed files with 284 additions and 0 deletions.
2 changes: 2 additions & 0 deletions tools/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,8 @@ git checkout tags/b3600
git apply ..\lcpp.patch
```

If you wish to use the SD3 version (`lcpp_sd3.patch`), which has slightly modified logic for quantizing, you'll want to target `tags/b3962` instead.


Then, compile the llama-quantize binary. This example uses cmake, on linux you can just use make.
```
Expand Down
282 changes: 282 additions & 0 deletions tools/lcpp_sd3.patch
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,282 @@
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index de3c706f..96bcab79 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -223,7 +223,7 @@
#define GGML_MAX_OP_PARAMS 64

#ifndef GGML_MAX_NAME
-# define GGML_MAX_NAME 64
+# define GGML_MAX_NAME 128
#endif

#define GGML_DEFAULT_N_THREADS 4
@@ -2449,6 +2449,7 @@ extern "C" {

// manage tensor info
GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor);
+ GGML_API void gguf_set_tensor_ndim(struct gguf_context * ctx, const char * name, uint32_t n_dim);
GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type);
GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size);

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index b16c462f..987cdcc2 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -22960,6 +22960,14 @@ void gguf_add_tensor(
ctx->header.n_tensors++;
}

+void gguf_set_tensor_ndim(struct gguf_context * ctx, const char * name, const uint32_t n_dim) {
+ const int idx = gguf_find_tensor(ctx, name);
+ if (idx < 0) {
+ GGML_ABORT("tensor not found");
+ }
+ ctx->infos[idx].n_dims = n_dim;
+}
+
void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type) {
const int idx = gguf_find_tensor(ctx, name);
if (idx < 0) {
diff --git a/src/llama.cpp b/src/llama.cpp
index 24e1f1f0..e7747711 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -205,6 +205,10 @@ enum llm_arch {
LLM_ARCH_GRANITE,
LLM_ARCH_GRANITE_MOE,
LLM_ARCH_CHAMELEON,
+ LLM_ARCH_FLUX,
+ LLM_ARCH_SD1,
+ LLM_ARCH_SDXL,
+ LLM_ARCH_SD3,
LLM_ARCH_UNKNOWN,
};

@@ -258,6 +262,10 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_GRANITE, "granite" },
{ LLM_ARCH_GRANITE_MOE, "granitemoe" },
{ LLM_ARCH_CHAMELEON, "chameleon" },
+ { LLM_ARCH_FLUX, "flux" },
+ { LLM_ARCH_SD1, "sd1" },
+ { LLM_ARCH_SDXL, "sdxl" },
+ { LLM_ARCH_SD3, "sd3" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};

@@ -1531,6 +1539,10 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
},
},
+ { LLM_ARCH_FLUX, {}},
+ { LLM_ARCH_SD1, {}},
+ { LLM_ARCH_SDXL, {}},
+ { LLM_ARCH_SD3, {}},
{
LLM_ARCH_UNKNOWN,
{
@@ -5403,6 +5415,12 @@ static void llm_load_hparams(
// get general kv
ml.get_key(LLM_KV_GENERAL_NAME, model.name, false);

+ // Disable LLM metadata for image models
+ if (model.arch == LLM_ARCH_FLUX || model.arch == LLM_ARCH_SD1 || model.arch == LLM_ARCH_SDXL || model.arch == LLM_ARCH_SD3) {
+ model.ftype = ml.ftype;
+ return;
+ }
+
// get hparams kv
ml.get_key(LLM_KV_VOCAB_SIZE, hparams.n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab);

@@ -18016,6 +18034,122 @@ static void llama_tensor_dequantize_internal(
workers.clear();
}

+static ggml_type img_tensor_get_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
+ // Special function for quantizing image model tensors
+ const std::string name = ggml_get_name(tensor);
+ const llm_arch arch = qs.model.arch;
+
+ // Sanity check
+ if (
+ (name.find("model.diffusion_model.") != std::string::npos) ||
+ (name.find("first_stage_model.") != std::string::npos) ||
+ (name.find("single_transformer_blocks.") != std::string::npos)
+ ) {
+ throw std::runtime_error("Invalid input GGUF file. This is not a supported UNET model");
+ }
+
+ // Unsupported quant types - exclude all IQ quants for now
+ if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS ||
+ ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ||
+ ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
+ ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL ||
+ ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||
+ ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_Q4_0_4_4 ||
+ ftype == LLAMA_FTYPE_MOSTLY_Q4_0_4_8 || ftype == LLAMA_FTYPE_MOSTLY_Q4_0_8_8) {
+ throw std::runtime_error("Invalid quantization type for image model (Not supported)");
+ }
+
+ if ( // Rules for to_v attention
+ (name.find("attn_v.weight") != std::string::npos) ||
+ (name.find(".to_v.weight") != std::string::npos)
+ ){
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
+ new_type = GGML_TYPE_Q3_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+ new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+ new_type = GGML_TYPE_Q5_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) {
+ new_type = GGML_TYPE_Q6_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) {
+ new_type = GGML_TYPE_Q5_K;
+ }
+ ++qs.i_attention_wv;
+ } else if ( // Rules for fused qkv attention
+ (name.find("attn_qkv.weight") != std::string::npos) ||
+ (name.find("attn.qkv.weight") != std::string::npos)
+ ) {
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+ new_type = GGML_TYPE_Q4_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+ new_type = GGML_TYPE_Q5_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) {
+ new_type = GGML_TYPE_Q6_K;
+ }
+ } else if ( // Rules for ffn
+ (name.find("ffn_down") != std::string::npos)
+ ) {
+ // TODO: add back `layer_info` with some model specific logic + logic further down
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
+ new_type = GGML_TYPE_Q4_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
+ new_type = GGML_TYPE_Q5_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S) {
+ new_type = GGML_TYPE_Q5_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+ new_type = GGML_TYPE_Q6_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) {
+ new_type = GGML_TYPE_Q6_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_0) {
+ new_type = GGML_TYPE_Q4_1;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_0) {
+ new_type = GGML_TYPE_Q5_1;
+ }
+ ++qs.i_ffn_down;
+ }
+
+ // Sanity check for row shape
+ bool convert_incompatible_tensor = false;
+ if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
+ new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
+ int nx = tensor->ne[0];
+ int ny = tensor->ne[1];
+ if (nx % QK_K != 0) {
+ LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
+ convert_incompatible_tensor = true;
+ } else {
+ ++qs.n_k_quantized;
+ }
+ }
+ if (convert_incompatible_tensor) {
+ // TODO: Possibly reenable this in the future
+ // switch (new_type) {
+ // case GGML_TYPE_Q2_K:
+ // case GGML_TYPE_Q3_K:
+ // case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
+ // case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q5_1; break;
+ // case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break;
+ // default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
+ // }
+ new_type = GGML_TYPE_F16;
+ LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
+ ++qs.n_fallback;
+ }
+ return new_type;
+}
+
static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
const std::string name = ggml_get_name(tensor);

@@ -18647,6 +18781,50 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// do not quantize relative position bias (T5)
quantize &= name.find("attn_rel_b.weight") == std::string::npos;

+ // rules for image models
+ bool image_model = false;
+ if (model.arch == LLM_ARCH_FLUX) {
+ image_model = true;
+ quantize &= name.find("txt_in.") == std::string::npos;
+ quantize &= name.find("img_in.") == std::string::npos;
+ quantize &= name.find("time_in.") == std::string::npos;
+ quantize &= name.find("vector_in.") == std::string::npos;
+ quantize &= name.find("guidance_in.") == std::string::npos;
+ quantize &= name.find("final_layer.") == std::string::npos;
+ }
+ if (model.arch == LLM_ARCH_SD1 || model.arch == LLM_ARCH_SDXL) {
+ image_model = true;
+ quantize &= name.find("class_embedding.") == std::string::npos;
+ quantize &= name.find("time_embedding.") == std::string::npos;
+ quantize &= name.find("add_embedding.") == std::string::npos;
+ quantize &= name.find("time_embed.") == std::string::npos;
+ quantize &= name.find("label_emb.") == std::string::npos;
+ quantize &= name.find("conv_in.") == std::string::npos;
+ quantize &= name.find("conv_out.") == std::string::npos;
+ quantize &= name != "input_blocks.0.0.weight";
+ quantize &= name != "out.2.weight";
+ }
+ if (model.arch == LLM_ARCH_SD3) {
+ image_model = true;
+ quantize &= name.find("final_layer.") == std::string::npos;
+ quantize &= name.find("time_text_embed.") == std::string::npos;
+ quantize &= name.find("context_embedder.") == std::string::npos;
+ quantize &= name.find("t_embedder.") == std::string::npos;
+ quantize &= name.find("y_embedder.") == std::string::npos;
+ quantize &= name.find("x_embedder.") == std::string::npos;
+ quantize &= name != "proj_out.weight";
+ quantize &= name != "pos_embed";
+ // SD3 pos_embed needs special fix as first dim is 1, which gets truncated here
+ if (name == "pos_embed" && tensor->ne[2] == 1) {
+ const uint32_t n_dim = 3;
+ gguf_set_tensor_ndim(ctx_outs[cur_split], "pos_embed", n_dim);
+ }
+ }
+ // ignore 3D/4D tensors for image models as the code was never meant to handle these
+ if (image_model) {
+ quantize &= ggml_n_dims(tensor) == 2;
+ }
+
enum ggml_type new_type;
void * new_data;
size_t new_size;
@@ -18655,6 +18833,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
new_type = default_type;

// get more optimal quantization type based on the tensor shape, layer, etc.
+ if (image_model) {
+ new_type = img_tensor_get_type(qs, new_type, tensor, ftype);
+ } else {
if (!params->pure && ggml_is_quantized(default_type)) {
new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
}
@@ -18664,6 +18845,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) {
new_type = params->output_tensor_type;
}
+ }

// If we've decided to quantize to the same type the tensor is already
// in then there's nothing to do.

0 comments on commit 4d7bb93

Please sign in to comment.