tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

scripts: n_depth for compare-llama-bench [no ci] (#13201)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-04-29 23:32:04 +02:00
committed by GitHub
parent e2e1ddb93a
commit 19e899ce21
1 changed files with 18 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
scripts/compare-llama-bench.py
View File

@ -19,9 +19,9 @@ logger = logging.getLogger("compare-llama-bench")
# Properties by which to differentiate results per commit:
KEY_PROPERTIES = [
"cpu_info", "gpu_info", "backends", "n_gpu_layers", "model_filename", "model_type", "n_batch", "n_ubatch",
"embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v", "use_mmap", "no_kv_offload",
"split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen"
"cpu_info", "gpu_info", "backends", "n_gpu_layers", "tensor_buft_overrides", "model_filename", "model_type",
"n_batch", "n_ubatch", "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v",
"use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen", "n_depth"
]
# Properties that are boolean and are converted to Yes/No for the table:
@ -30,11 +30,11 @@ BOOL_PROPERTIES = ["embeddings", "cpu_strict", "use_mmap", "no_kv_offload", "fla
# Header names for the table:
PRETTY_NAMES = {
"cpu_info": "CPU", "gpu_info": "GPU", "backends": "Backends", "n_gpu_layers": "GPU layers",
"model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
"model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size",
"embeddings": "Embeddings", "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll",
"n_threads": "Threads", "type_k": "K type", "type_v": "V type", "split_mode": "Split mode", "main_gpu": "Main GPU",
"no_kv_offload": "NKVO", "flash_attn": "FlashAttention", "tensor_split": "Tensor split", "use_mmap": "Use mmap",
"tensor_buft_overrides": "Tensor overrides", "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
"model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size", "embeddings": "Embeddings",
"cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll", "n_threads": "Threads", "type_k": "K type", "type_v": "V type",
"use_mmap": "Use mmap", "no_kv_offload": "NKVO", "split_mode": "Split mode", "main_gpu": "Main GPU", "tensor_split": "Tensor split",
"flash_attn": "FlashAttention",
}
DEFAULT_SHOW = ["model_type"] # Always show these properties by default.
@ -281,12 +281,12 @@ def get_rows(properties):
The returned rows are unique in terms of property combinations.
"""
select_string = ", ".join(
[f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
[f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "tb.n_depth", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
equal_string = " AND ".join(
[f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
)
group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt"])
group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt", "tb.n_depth"])
query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
return cursor.execute(query).fetchall()
@ -309,7 +309,7 @@ else:
rows_full = get_rows(KEY_PROPERTIES)
properties_different = []
for i, kp_i in enumerate(KEY_PROPERTIES):
if kp_i in DEFAULT_SHOW or kp_i == "n_prompt" or kp_i == "n_gen":
if kp_i in DEFAULT_SHOW or kp_i in ["n_prompt", "n_gen", "n_depth"]:
continue
for row_full in rows_full:
if row_full[i] != rows_full[0][i]:
@ -340,17 +340,20 @@ else:
table = []
for row in rows_show:
n_prompt = int(row[-4])
n_gen = int(row[-3])
n_prompt = int(row[-5])
n_gen = int(row[-4])
n_depth = int(row[-3])
if n_prompt != 0 and n_gen == 0:
test_name = f"pp{n_prompt}"
elif n_prompt == 0 and n_gen != 0:
test_name = f"tg{n_gen}"
else:
test_name = f"pp{n_prompt}+tg{n_gen}"
if n_depth != 0:
test_name = f"{test_name}@d{n_depth}"
# Regular columns test name avg t/s values Speedup
# VVVVVVVVVVVVV VVVVVVVVV VVVVVVVVVVVVVV VVVVVVV
table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
table.append(list(row[:-5]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
# Some a-posteriori fixes to make the table contents prettier:
for bool_property in BOOL_PROPERTIES:
@ -376,7 +379,7 @@ if "gpu_info" in show:
for gns in GPU_NAME_STRIP:
row_table[ip] = row_table[ip].replace(gns, "")
gpu_names = row_table[ip].split("/")
gpu_names = row_table[ip].split(", ")
num_gpus = len(gpu_names)
all_names_the_same = len(set(gpu_names)) == 1
if len(gpu_names) >= 2 and all_names_the_same: