Skip to content

Commit

Permalink
update custom reporting script
Browse files Browse the repository at this point in the history
  • Loading branch information
@IlyasMoutawwakil
IlyasMoutawwakil committed Nov 27, 2023
1 parent 8aec23b commit 225d165
Showing 1 changed file with 60 additions and 85 deletions.
145 changes: 60 additions & 85 deletions examples/running-llamas/report.py → examples/running-llamas/batch_size_report.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -49,27 +49,30 @@ def get_short_report(full_report, report_folder: str = "artifacts"):
"model": "Model",
"environment.gpus": "GPUs",
"experiment_name": "Experiment Name",
"benchmark.input_shapes.batch_size": "Batch Size",
"benchmark.input_shapes.batch_size": "Per Process Batch Size",
"benchmark.input_shapes.sequence_length": "Sequence Length",
#
"decode.latency(s)": "Decode Latency (s)",
"forward.latency(s)": "Prefill Latency (s)",
"generate.latency(s)": "Generate Latency (s)",
#
"decode.throughput(tokens/s)": "Decode Throughput (tokens/s)",
"forward.throughput(samples/s)": "Prefill Throughput (samples/s)",
"generate.throughput(tokens/s)": "Generate Throughput (tokens/s)",
#
"forward.max_memory_allocated(MB)": "Prefill Memory (MB)",
"generate.max_memory_allocated(MB)": "Decode Memory (MB)",
}
short_report = full_report[list(short_columns.keys())].rename(columns=short_columns)

short_report["GPU Name"] = short_report["GPUs"].str[0]
short_report["Num GPUs"] = short_report["GPUs"].str.len().astype(str)
short_report["Num GPUs"] = short_report["GPUs"].str.len()
short_report["GPU Name"].replace("NVIDIA A100-SXM4-80GB", "A100", inplace=True)
short_report["GPU Name"].replace("AMD INSTINCT MI250 (MCM) OAM AC MBA", "GCD-MI250", inplace=True)
short_report["Effective Batch Size"] = short_report["Per Process Batch Size"] * short_report["Num GPUs"]

short_report["Group"] = (
short_report["Num GPUs"] + "x" + short_report["GPU Name"] + "-" + short_report["Experiment Name"]
short_report["Num GPUs"].astype(str) + "x" + short_report["GPU Name"] + "-" + short_report["Experiment Name"]
)

short_report.to_csv(f"{report_folder}/short_report.csv")

return short_report
Expand All @@ -78,123 +81,91 @@ def get_short_report(full_report, report_folder: str = "artifacts"):
import numpy as np


def get_plots(short_report, report_folder: str = "artifacts", plot="bar"):
def get_plots(short_report, report_folder, batch="effective", plot="bar"):
fig1, ax1 = plt.subplots()
fig2, ax2 = plt.subplots()
fig3, ax3 = plt.subplots()
fig4, ax4 = plt.subplots()

batch_column = "Effective Batch Size" if batch == "effective" else "Per Process Batch Size"

groups = sorted(short_report["Group"].unique().tolist())
x = np.arange(len(short_report["Batch Size"].unique()))
multiplier = 0
width = 0.2
x = np.arange(len(short_report[batch_column].unique()))
width = 0.8 / len(groups)
offset = -(width * (len(groups) - 1) / 2)

for group in groups:
offset = width * multiplier
mask = short_report["Group"] == group
group_report = short_report[mask].sort_values(by="Batch Size")
group_report = short_report[mask].sort_values(by=batch_column)

prefill_latency = group_report[["Batch Size", "Prefill Latency (s)"]]
decode_throughput = group_report[["Batch Size", "Decode Throughput (tokens/s)"]]
#
generate_latency = group_report[["Batch Size", "Generate Latency (s)"]]
generate_throughput = group_report[["Batch Size", "Generate Throughput (tokens/s)"]]
prefill_latency = group_report[[batch_column, "Prefill Latency (s)"]]
decode_throughput = group_report[[batch_column, "Decode Throughput (tokens/s)"]]

if plot == "bar":
x_ = np.arange(
prefill_latency[batch_column].min() - 1,
len(prefill_latency[batch_column].unique()) + (prefill_latency[batch_column].min() - 1),
)
ax1.bar(
(x + offset)[: len(prefill_latency["Batch Size"])],
x_ + offset,
prefill_latency["Prefill Latency (s)"],
label=group,
width=width,
)
x_ = np.arange(
decode_throughput[batch_column].min() - 1,
len(decode_throughput[batch_column].unique()) + (decode_throughput[batch_column].min() - 1),
)
ax2.bar(
(x + offset)[: len(decode_throughput["Batch Size"])],
x_ + offset,
decode_throughput["Decode Throughput (tokens/s)"],
label=group,
width=width,
)
#
ax3.bar(
(x + offset)[: len(generate_latency["Batch Size"])],
generate_latency["Generate Latency (s)"],
label=group,
width=width,
)
ax4.bar(
(x + offset)[: len(generate_throughput["Batch Size"])],
generate_throughput["Generate Throughput (tokens/s)"],
label=group,
width=width,
)
offset += width
elif plot == "line":
x_ = np.arange(
prefill_latency[batch_column].min() - 1,
len(prefill_latency[batch_column].unique()) + (prefill_latency[batch_column].min() - 1),
)
ax1.plot(
prefill_latency["Batch Size"],
x_,
prefill_latency["Prefill Latency (s)"],
label=group,
marker="o",
)
x_ = np.arange(
decode_throughput[batch_column].min() - 1,
len(decode_throughput[batch_column].unique()) + (decode_throughput[batch_column].min() - 1),
)
ax2.plot(
decode_throughput["Batch Size"],
x_,
decode_throughput["Decode Throughput (tokens/s)"],
label=group,
marker="o",
)
#
ax3.plot(
generate_latency["Batch Size"],
generate_latency["Generate Latency (s)"],
label=group,
marker="o",
)
ax4.plot(
generate_throughput["Batch Size"],
generate_throughput["Generate Throughput (tokens/s)"],
label=group,
marker="o",
)

multiplier += 1

if plot == "bar":
ax1.set_xticks(x)
ax1.set_xticklabels(short_report["Batch Size"].sort_values().unique().tolist())
ax1.set_xticks(x)
ax1.set_ylim(bottom=0)
ax1.set_xticklabels(short_report[batch_column].sort_values().unique().tolist())

ax2.set_xticks(x)
ax2.set_xticklabels(short_report["Batch Size"].sort_values().unique().tolist())
ax2.set_xticks(x)
ax2.set_ylim(bottom=0)
ax2.set_xticklabels(short_report[batch_column].sort_values().unique().tolist())

ax3.set_xticks(x)
ax3.set_xticklabels(short_report["Batch Size"].sort_values().unique().tolist())

ax4.set_xticks(x)
ax4.set_xticklabels(short_report["Batch Size"].sort_values().unique().tolist())

ax1.set_xlabel("Batch Size")
ax1.set_xlabel(batch_column)
ax1.set_ylabel("Prefill Latency (s)")
ax1.set_title(f"{short_report['Model'].unique()[0]} Prefill Latency per Batch Size")

ax2.set_xlabel("Batch Size")
ax2.set_ylabel("Decode Throughput (tokens/s)")
ax2.set_title(f"{short_report['Model'].unique()[0]} Decode Throughput per Batch Size")

ax3.set_xlabel("Batch Size")
ax3.set_ylabel("Generate Latency (s)")
ax3.set_title(f"{short_report['Model'].unique()[0]} Generate Latency per Batch Size")
ax1.set_title(f"Prefill Latency per Batch Size ({short_report['Model'].unique()[0]})")

ax4.set_xlabel("Batch Size")
ax4.set_ylabel("Generate Throughput (tokens/s)")
ax4.set_title(f"{short_report['Model'].unique()[0]} Generate Throughput per Batch Size")
ax2.set_xlabel(batch_column)
ax2.set_ylabel("Effective Decode Throughput (tokens/s)")
ax2.set_title(f"Decode Throughput per Batch Size ({short_report['Model'].unique()[0]})")

ax1.legend(fancybox=True, shadow=True)
ax2.legend(fancybox=True, shadow=True)
ax3.legend(fancybox=True, shadow=True)
ax4.legend(fancybox=True, shadow=True)

fig1.savefig(f"{report_folder}/prefill_latency_plot.png")
fig2.savefig(f"{report_folder}/decode_throughput_plot.png")
# fig3.savefig(f"{report_folder}/generate_latency_plot.png")
# fig4.savefig(f"{report_folder}/generate_throughput_plot.png")
fig1.savefig(f"{report_folder}/prefill_latency_{plot}_plot_{batch}.png")
fig2.savefig(f"{report_folder}/decode_throughput_{plot}_plot_{batch}.png")

return fig1, fig2, fig3, fig4
return fig1, fig2


def generate_report():
Expand Down Expand Up @@ -231,10 +202,14 @@ def generate_report():
full_report,
report_folder=report_folder,
)
figs = get_plots(
short_report,
report_folder=report_folder,
)
for plot in ["bar", "line"]:
for batch in ["effective", "per-process"]:
figs = get_plots(
short_report,
report_folder,
batch=batch,
plot=plot,
)


if __name__ == "__main__":
Expand Down

0 comments on commit 225d165

Please sign in to comment.