Skip to content

Commit fea9018

Browse files
mudlermudler
authored
Revert "feat(stablediffusion): Passthrough more parameters to support z-image and flux2" (#7417)
Revert "feat(stablediffusion): Passthrough more parameters to support z-image…" This reverts commit 4018e59.
1 parent d8c7e90 commit fea9018

2 files changed

Lines changed: 8 additions & 237 deletions

File tree

‎backend/go/stablediffusion-ggml/gosd.cpp‎

Lines changed: 8 additions & 234 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,4 @@
11
#include "stable-diffusion.h"
2-
#include <cmath>
32
#include <cstdint>
43
#define GGML_MAX_NAME 128
54

@@ -22,7 +21,6 @@
2221
#define STB_IMAGE_RESIZE_IMPLEMENTATION
2322
#define STB_IMAGE_RESIZE_STATIC
2423
#include "stb_image_resize.h"
25-
#include <stdlib.h>
2624

2725
// Names of the sampler method, same order as enum sample_method in stable-diffusion.h
2826
const char* sample_method_str[] = {
@@ -57,68 +55,6 @@ const char* schedulers[] = {
5755

5856
static_assert(std::size(schedulers) == SCHEDULER_COUNT, "schedulers mismatch");
5957

60-
// New enum string arrays
61-
const char* rng_type_str[] = {
62-
"std_default",
63-
"cuda",
64-
"cpu",
65-
};
66-
static_assert(std::size(rng_type_str) == RNG_TYPE_COUNT, "rng type mismatch");
67-
68-
const char* prediction_str[] = {
69-
"default",
70-
"epsilon",
71-
"v",
72-
"edm_v",
73-
"sd3_flow",
74-
"flux_flow",
75-
"flux2_flow",
76-
};
77-
static_assert(std::size(prediction_str) == PREDICTION_COUNT, "prediction mismatch");
78-
79-
const char* lora_apply_mode_str[] = {
80-
"auto",
81-
"immediately",
82-
"at_runtime",
83-
};
84-
static_assert(std::size(lora_apply_mode_str) == LORA_APPLY_MODE_COUNT, "lora apply mode mismatch");
85-
86-
const char* sd_type_str[] = {
87-
[0] = "f32",
88-
[1] = "f16",
89-
[2] = "q4_0",
90-
[3] = "q4_1",
91-
[6] = "q5_0",
92-
[7] = "q5_1",
93-
[8] = "q8_0",
94-
[9] = "q8_1",
95-
[10] = "q2_k",
96-
[11] = "q3_k",
97-
[12] = "q4_k",
98-
[13] = "q5_k",
99-
[14] = "q6_k",
100-
[15] = "q8_k",
101-
[16] = "iq2_xxs",
102-
[17] = "iq2_xs",
103-
[18] = "iq3_xxs",
104-
[19] = "iq1_s",
105-
[20] = "iq4_nl",
106-
[21] = "iq3_s",
107-
[22] = "iq2_s",
108-
[23] = "iq4_xs",
109-
[24] = "i8",
110-
[25] = "i16",
111-
[26] = "i32",
112-
[27] = "i64",
113-
[28] = "f64",
114-
[29] = "iq1_m",
115-
[30] = "bf16",
116-
[34] = "tq1_0",
117-
[35] = "tq2_0",
118-
[39] = "mxfp4",
119-
};
120-
static_assert(std::size(sd_type_str) == SD_TYPE_COUNT, "sd type mismatch");
121-
12258
sd_ctx_t* sd_c;
12359
// Moved from the context (load time) to generation time params
12460
scheduler_t scheduler = SCHEDULER_COUNT;
@@ -174,41 +110,9 @@ int load_model(const char *model, char *model_path, char* options[], int threads
174110
const char *vae_path = "";
175111
const char *scheduler_str = "";
176112
const char *sampler = "";
177-
const char *clip_vision_path = "";
178-
const char *llm_path = "";
179-
const char *llm_vision_path = "";
180-
const char *diffusion_model_path = stableDiffusionModel;
181-
const char *high_noise_diffusion_model_path = "";
182-
const char *taesd_path = "";
183-
const char *control_net_path = "";
184-
const char *embedding_dir = "";
185-
const char *photo_maker_path = "";
186-
const char *tensor_type_rules = "";
187113
char *lora_dir = model_path;
188114
bool lora_dir_allocated = false;
189115

190-
bool vae_decode_only = true;
191-
bool free_params_immediately = true;
192-
int n_threads = threads;
193-
enum sd_type_t wtype = SD_TYPE_COUNT;
194-
enum rng_type_t rng_type = STD_DEFAULT_RNG;
195-
enum rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
196-
enum prediction_t prediction = PREDICTION_COUNT;
197-
enum lora_apply_mode_t lora_apply_mode = LORA_APPLY_MODE_COUNT;
198-
bool offload_params_to_cpu = false;
199-
bool keep_clip_on_cpu = false;
200-
bool keep_control_net_on_cpu = false;
201-
bool keep_vae_on_cpu = false;
202-
bool diffusion_flash_attn = false;
203-
bool tae_preview_only = false;
204-
bool diffusion_conv_direct = false;
205-
bool vae_conv_direct = false;
206-
bool force_sdxl_vae_conv_scale = false;
207-
bool chroma_use_dit_mask = true;
208-
bool chroma_use_t5_mask = false;
209-
int chroma_t5_mask_pad = 0;
210-
float flow_shift = INFINITY;
211-
212116
fprintf(stderr, "parsing options: %p\n", options);
213117

214118
// If options is not NULL, parse options
@@ -252,113 +156,6 @@ int load_model(const char *model, char *model_path, char* options[], int threads
252156
fprintf(stderr, "No model path provided, using lora dir as-is: %s\n", lora_dir);
253157
}
254158
}
255-
256-
// New parsing
257-
if (!strcmp(optname, "clip_vision_path")) clip_vision_path = optval;
258-
if (!strcmp(optname, "llm_path")) llm_path = optval;
259-
if (!strcmp(optname, "llm_vision_path")) llm_vision_path = optval;
260-
if (!strcmp(optname, "diffusion_model_path")) diffusion_model_path = optval;
261-
if (!strcmp(optname, "high_noise_diffusion_model_path")) high_noise_diffusion_model_path = optval;
262-
if (!strcmp(optname, "taesd_path")) taesd_path = optval;
263-
if (!strcmp(optname, "control_net_path")) control_net_path = optval;
264-
if (!strcmp(optname, "embedding_dir")) embedding_dir = optval;
265-
if (!strcmp(optname, "photo_maker_path")) photo_maker_path = optval;
266-
if (!strcmp(optname, "tensor_type_rules")) tensor_type_rules = optval;
267-
268-
if (!strcmp(optname, "vae_decode_only")) vae_decode_only = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
269-
if (!strcmp(optname, "free_params_immediately")) free_params_immediately = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
270-
if (!strcmp(optname, "offload_params_to_cpu")) offload_params_to_cpu = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
271-
if (!strcmp(optname, "keep_clip_on_cpu")) keep_clip_on_cpu = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
272-
if (!strcmp(optname, "keep_control_net_on_cpu")) keep_control_net_on_cpu = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
273-
if (!strcmp(optname, "keep_vae_on_cpu")) keep_vae_on_cpu = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
274-
if (!strcmp(optname, "diffusion_flash_attn")) diffusion_flash_attn = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
275-
if (!strcmp(optname, "tae_preview_only")) tae_preview_only = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
276-
if (!strcmp(optname, "diffusion_conv_direct")) diffusion_conv_direct = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
277-
if (!strcmp(optname, "vae_conv_direct")) vae_conv_direct = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
278-
if (!strcmp(optname, "force_sdxl_vae_conv_scale")) force_sdxl_vae_conv_scale = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
279-
if (!strcmp(optname, "chroma_use_dit_mask")) chroma_use_dit_mask = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
280-
if (!strcmp(optname, "chroma_use_t5_mask")) chroma_use_t5_mask = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
281-
282-
if (!strcmp(optname, "n_threads")) n_threads = atoi(optval);
283-
if (!strcmp(optname, "chroma_t5_mask_pad")) chroma_t5_mask_pad = atoi(optval);
284-
285-
if (!strcmp(optname, "flow_shift")) flow_shift = atof(optval);
286-
287-
if (!strcmp(optname, "rng_type")) {
288-
int found = -1;
289-
for (int m = 0; m < RNG_TYPE_COUNT; m++) {
290-
if (!strcmp(optval, rng_type_str[m])) {
291-
found = m;
292-
break;
293-
}
294-
}
295-
if (found != -1) {
296-
rng_type = (rng_type_t)found;
297-
fprintf(stderr, "Found rng_type: %s\n", optval);
298-
} else {
299-
fprintf(stderr, "Invalid rng_type: %s, using default\n", optval);
300-
}
301-
}
302-
if (!strcmp(optname, "sampler_rng_type")) {
303-
int found = -1;
304-
for (int m = 0; m < RNG_TYPE_COUNT; m++) {
305-
if (!strcmp(optval, rng_type_str[m])) {
306-
found = m;
307-
break;
308-
}
309-
}
310-
if (found != -1) {
311-
sampler_rng_type = (rng_type_t)found;
312-
fprintf(stderr, "Found sampler_rng_type: %s\n", optval);
313-
} else {
314-
fprintf(stderr, "Invalid sampler_rng_type: %s, using default\n", optval);
315-
}
316-
}
317-
if (!strcmp(optname, "prediction")) {
318-
int found = -1;
319-
for (int m = 0; m < PREDICTION_COUNT; m++) {
320-
if (!strcmp(optval, prediction_str[m])) {
321-
found = m;
322-
break;
323-
}
324-
}
325-
if (found != -1) {
326-
prediction = (prediction_t)found;
327-
fprintf(stderr, "Found prediction: %s\n", optval);
328-
} else {
329-
fprintf(stderr, "Invalid prediction: %s, using default\n", optval);
330-
}
331-
}
332-
if (!strcmp(optname, "lora_apply_mode")) {
333-
int found = -1;
334-
for (int m = 0; m < LORA_APPLY_MODE_COUNT; m++) {
335-
if (!strcmp(optval, lora_apply_mode_str[m])) {
336-
found = m;
337-
break;
338-
}
339-
}
340-
if (found != -1) {
341-
lora_apply_mode = (lora_apply_mode_t)found;
342-
fprintf(stderr, "Found lora_apply_mode: %s\n", optval);
343-
} else {
344-
fprintf(stderr, "Invalid lora_apply_mode: %s, using default\n", optval);
345-
}
346-
}
347-
if (!strcmp(optname, "wtype")) {
348-
int found = -1;
349-
for (int m = 0; m < SD_TYPE_COUNT; m++) {
350-
if (sd_type_str[m] && !strcmp(optval, sd_type_str[m])) {
351-
found = m;
352-
break;
353-
}
354-
}
355-
if (found != -1) {
356-
wtype = (sd_type_t)found;
357-
fprintf(stderr, "Found wtype: %s\n", optval);
358-
} else {
359-
fprintf(stderr, "Invalid wtype: %s, using default\n", optval);
360-
}
361-
}
362159
}
363160

364161
fprintf(stderr, "parsed options\n");
@@ -369,40 +166,17 @@ int load_model(const char *model, char *model_path, char* options[], int threads
369166
ctx_params.model_path = model;
370167
ctx_params.clip_l_path = clip_l_path;
371168
ctx_params.clip_g_path = clip_g_path;
372-
ctx_params.clip_vision_path = clip_vision_path;
373169
ctx_params.t5xxl_path = t5xxl_path;
374-
ctx_params.llm_path = llm_path;
375-
ctx_params.llm_vision_path = llm_vision_path;
376-
ctx_params.diffusion_model_path = diffusion_model_path;
377-
ctx_params.high_noise_diffusion_model_path = high_noise_diffusion_model_path;
170+
ctx_params.diffusion_model_path = stableDiffusionModel;
378171
ctx_params.vae_path = vae_path;
379-
ctx_params.taesd_path = taesd_path;
380-
ctx_params.control_net_path = control_net_path;
172+
ctx_params.taesd_path = "";
173+
ctx_params.control_net_path = "";
381174
ctx_params.lora_model_dir = lora_dir;
382-
ctx_params.embedding_dir = embedding_dir;
383-
ctx_params.photo_maker_path = photo_maker_path;
384-
ctx_params.tensor_type_rules = tensor_type_rules;
385-
ctx_params.vae_decode_only = vae_decode_only;
386-
ctx_params.free_params_immediately = free_params_immediately;
387-
ctx_params.n_threads = n_threads;
388-
ctx_params.rng_type = rng_type;
389-
ctx_params.keep_clip_on_cpu = keep_clip_on_cpu;
390-
if (wtype != SD_TYPE_COUNT) ctx_params.wtype = wtype;
391-
if (sampler_rng_type != RNG_TYPE_COUNT) ctx_params.sampler_rng_type = sampler_rng_type;
392-
if (prediction != PREDICTION_COUNT) ctx_params.prediction = prediction;
393-
if (lora_apply_mode != LORA_APPLY_MODE_COUNT) ctx_params.lora_apply_mode = lora_apply_mode;
394-
ctx_params.offload_params_to_cpu = offload_params_to_cpu;
395-
ctx_params.keep_control_net_on_cpu = keep_control_net_on_cpu;
396-
ctx_params.keep_vae_on_cpu = keep_vae_on_cpu;
397-
ctx_params.diffusion_flash_attn = diffusion_flash_attn;
398-
ctx_params.tae_preview_only = tae_preview_only;
399-
ctx_params.diffusion_conv_direct = diffusion_conv_direct;
400-
ctx_params.vae_conv_direct = vae_conv_direct;
401-
ctx_params.force_sdxl_vae_conv_scale = force_sdxl_vae_conv_scale;
402-
ctx_params.chroma_use_dit_mask = chroma_use_dit_mask;
403-
ctx_params.chroma_use_t5_mask = chroma_use_t5_mask;
404-
ctx_params.chroma_t5_mask_pad = chroma_t5_mask_pad;
405-
ctx_params.flow_shift = flow_shift;
175+
ctx_params.embedding_dir = "";
176+
ctx_params.vae_decode_only = false;
177+
ctx_params.free_params_immediately = false;
178+
ctx_params.n_threads = threads;
179+
ctx_params.rng_type = STD_DEFAULT_RNG;
406180
sd_ctx_t* sd_ctx = new_sd_ctx(&ctx_params);
407181

408182
if (sd_ctx == NULL) {

‎gallery/index.yaml‎

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -20911,9 +20911,6 @@
2091120911
overrides:
2091220912
parameters:
2091320913
model: flux1-dev-Q2_K.gguf
20914-
options:
20915-
- scheduler:simple
20916-
- keep_clip_on_cpu:true
2091720914
files:
2091820915
- filename: "flux1-dev-Q2_K.gguf"
2091920916
sha256: "b8c464bc0f10076ef8f00ba040d220d90c7993f7c4245ae80227d857f65df105"

0 commit comments

Comments
 (0)