olmocr/tests/test_sglang.py

401 lines
16 KiB
Python
Raw Normal View History

2024-11-25 09:34:59 -08:00
# The idea is that you have a Qwen2-VL-7B model located here:s3://ai2-oe-data/jakep/experiments/qwen2vl-pdf/v1/models/jakep/Qwen_Qwen2-VL-7B-Instruct-e4ecf8-01JAH8GMWHTJ376S2N7ETXRXH4/checkpoint-9500/bf16/"
# You need to load it in both hugging face transformers, and send page 1 of edgar.pdf to it from tests/gnarly_pdfs
# Compare that the temperature 0 sampled result is the same
import asyncio
2025-01-29 15:25:10 -08:00
import base64
2024-11-25 09:34:59 -08:00
import json
2024-11-25 11:00:03 -08:00
import math
2025-01-29 15:25:10 -08:00
import os
import unittest
2024-11-25 10:25:55 -08:00
from io import BytesIO
2024-11-25 09:34:59 -08:00
from pathlib import Path
2025-01-29 15:25:10 -08:00
from unittest.mock import AsyncMock, patch
import numpy as np
2025-02-14 20:42:19 +00:00
import pytest
2025-01-29 15:25:10 -08:00
import torch
2024-11-25 11:00:03 -08:00
import torch.nn.functional as F
2025-01-29 15:25:10 -08:00
from httpx import AsyncClient
from PIL import Image
from transformers import AutoProcessor, AutoTokenizer, Qwen2VLForConditionalGeneration
2025-01-29 15:15:10 -08:00
2025-01-29 15:25:10 -08:00
from olmocr.pipeline import (
SGLANG_SERVER_PORT,
build_page_query,
get_anchor_text,
render_pdf_to_base64png,
sglang_server_ready,
sglang_server_task,
)
from olmocr.prompts import PageResponse
2025-01-29 15:15:10 -08:00
2025-01-29 15:30:39 -08:00
MODEL_FINETUNED_PATH = (
"s3://ai2-oe-data/jakep/experiments/qwen2vl-pdf/v1/models/jakep/Qwen_Qwen2-VL-7B-Instruct-e4ecf8-01JAH8GMWHTJ376S2N7ETXRXH4/checkpoint-9500/bf16/"
)
2024-11-25 10:12:29 -08:00
2025-02-14 20:42:19 +00:00
@pytest.mark.nonci
2024-11-25 09:34:59 -08:00
class TestSglangServer(unittest.IsolatedAsyncioTestCase):
async def asyncSetUp(self):
# Mock arguments
self.args = AsyncMock()
self.args.workspace = "/tmp/test_workspace"
2024-11-25 10:12:29 -08:00
self.args.model = [MODEL_FINETUNED_PATH]
2024-11-25 09:34:59 -08:00
self.args.model_chat_template = "qwen2-vl"
self.args.target_longest_image_dim = 1024
self.args.target_anchor_text_len = 6000
self.args.model_max_context = 8192
# Create a temporary workspace directory
os.makedirs(self.args.workspace, exist_ok=True)
# Set up a semaphore for server tasks
self.semaphore = asyncio.Semaphore(1)
2024-11-25 11:00:03 -08:00
self.maxDiff = None
2024-11-25 09:34:59 -08:00
2024-11-25 11:24:21 -08:00
# # Start the sglang server
# self.my_server_task = asyncio.create_task(sglang_server_task(self.args, self.semaphore))
2025-01-29 15:30:39 -08:00
2024-11-25 11:24:21 -08:00
# # Wait for the server to become ready
# await sglang_server_ready()
2024-11-25 09:48:05 -08:00
2024-11-25 11:24:21 -08:00
async def test_sglang_server_initialization_and_request(self):
2024-11-25 09:48:05 -08:00
# Mock data paths
2024-12-04 13:20:10 -08:00
self.test_pdf_path = Path(os.path.join(os.path.dirname(__file__), "gnarly_pdfs", "ambiguous.pdf"))
2024-11-25 09:48:05 -08:00
2024-11-25 09:34:59 -08:00
# Send a single request to the sglang server for page 1
async with AsyncClient(timeout=600) as session:
query = await build_page_query(
str(self.test_pdf_path),
page=1,
target_longest_image_dim=self.args.target_longest_image_dim,
target_anchor_text_len=self.args.target_anchor_text_len,
)
2024-11-25 11:24:21 -08:00
COMPLETION_URL = f"http://localhost:{30000}/v1/chat/completions"
query["temperature"] = 0.0
query["logprobs"] = True
2024-11-25 15:36:04 -08:00
query["top_logprobs"] = 5
2024-11-25 09:34:59 -08:00
response = await session.post(COMPLETION_URL, json=query)
2024-11-25 11:24:21 -08:00
print(response.text)
2024-11-25 09:34:59 -08:00
# Check the server response
self.assertEqual(response.status_code, 200)
response_data = response.json()
self.assertIn("choices", response_data)
self.assertGreater(len(response_data["choices"]), 0)
2024-11-25 09:39:55 -08:00
model_response_json = json.loads(response_data["choices"][0]["message"]["content"])
page_response = PageResponse(**model_response_json)
print(page_response)
2024-11-25 10:12:29 -08:00
self.assertEqual(page_response.natural_text, EDGAR_TEXT)
2024-11-25 09:48:05 -08:00
async def asyncTearDown(self):
2024-11-25 11:24:21 -08:00
pass
# # Shut down the server
# self.my_server_task.cancel()
# with self.assertRaises(asyncio.CancelledError):
# await self.my_server_task
# # Cleanup temporary workspace
# if os.path.exists(self.args.workspace):
# for root, _, files in os.walk(self.args.workspace):
# for file in files:
# os.unlink(os.path.join(root, file))
# os.rmdir(self.args.workspace)
2024-11-25 10:12:29 -08:00
2025-02-14 20:42:19 +00:00
@pytest.mark.nonci
2024-11-25 10:12:29 -08:00
class TestHuggingFaceModel(unittest.IsolatedAsyncioTestCase):
async def asyncSetUp(self):
# Set up the Hugging Face model and tokenizer
2025-01-29 15:30:39 -08:00
model_cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "olmocr", "model")
2024-11-25 10:12:29 -08:00
download_directory([MODEL_FINETUNED_PATH], model_cache_dir)
# Check the rope config and make sure it's got the proper key
with open(os.path.join(model_cache_dir, "config.json"), "r") as cfin:
config_data = json.load(cfin)
if "rope_type" in config_data["rope_scaling"]:
del config_data["rope_scaling"]["rope_type"]
config_data["rope_scaling"]["type"] = "mrope"
with open(os.path.join(model_cache_dir, "config.json"), "w") as cfout:
json.dump(config_data, cfout)
self.tokenizer = AutoTokenizer.from_pretrained(model_cache_dir, trust_remote_code=True)
2024-12-03 15:32:53 -08:00
self.image_token_id = self.tokenizer.encode("<|image_pad|>")[0]
2024-11-25 10:12:29 -08:00
self.model = Qwen2VLForConditionalGeneration.from_pretrained(model_cache_dir, torch_dtype=torch.bfloat16, trust_remote_code=True).eval()
2024-11-25 10:25:55 -08:00
self.processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
2024-11-25 10:12:29 -08:00
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.model.to(self.device)
# Path to the test PDF
2024-12-04 13:20:10 -08:00
self.test_pdf_path = Path(os.path.join(os.path.dirname(__file__), "gnarly_pdfs", "ambiguous.pdf"))
2024-11-25 11:00:03 -08:00
self.maxDiff = None
2024-11-25 10:12:29 -08:00
async def test_hugging_face_generation(self):
2024-11-25 10:25:55 -08:00
query = await build_page_query(
2025-01-29 15:30:39 -08:00
str(self.test_pdf_path),
page=1,
target_longest_image_dim=1024,
target_anchor_text_len=6000,
)
2024-11-25 10:25:55 -08:00
2024-12-03 15:32:53 -08:00
messages = query["messages"]
2025-01-29 15:30:39 -08:00
# Apply chat template to get the text
text = self.processor.apply_chat_template(query["messages"], tokenize=False, add_generation_prompt=True)
2024-11-25 10:25:55 -08:00
image_url = query["messages"][0]["content"][1]["image_url"]["url"]
# Remove the "data:image/png;base64," prefix
base64_image = image_url.split(",")[1]
# Decode the base64 string into bytes
image_data = base64.b64decode(base64_image)
# Create a BytesIO object and load it into a PIL image
main_image = Image.open(BytesIO(image_data))
# Process inputs using processor
inputs = self.processor(
text=[text],
images=[main_image],
padding=True,
return_tensors="pt",
2024-11-25 10:12:29 -08:00
)
2025-01-29 15:30:39 -08:00
image_indices = [idx for idx, token in enumerate(inputs["input_ids"][0]) if token.item() == self.image_token_id]
2024-12-03 15:32:53 -08:00
print("IMAGE INDICES", image_indices)
2024-11-26 08:38:25 -08:00
print(f"image_grid_thw - {inputs['image_grid_thw'].shape} {inputs['image_grid_thw']}")
print(f"pixel_values - {inputs['pixel_values'].shape} {inputs['pixel_values'].detach().cpu().numpy()}")
2025-01-29 15:30:39 -08:00
np.save("/root/pixel_values.npy", inputs["pixel_values"].detach().cpu().numpy())
2024-11-26 08:38:25 -08:00
2024-11-25 10:25:55 -08:00
inputs = {key: value.to(self.device) for (key, value) in inputs.items()}
2024-11-25 10:12:29 -08:00
2024-11-25 15:36:04 -08:00
generated_tokens = []
2024-12-04 13:20:10 -08:00
max_steps = 50
2024-11-25 15:36:04 -08:00
top_logprobs_hf = []
for step in range(max_steps):
# Generate the output with temperature=0
generation_output = self.model.generate(
**inputs,
temperature=0.0,
max_new_tokens=1,
2025-01-29 15:30:39 -08:00
# max_length=8192,
2024-11-25 15:36:04 -08:00
num_return_sequences=1,
do_sample=False,
output_scores=True,
return_dict_in_generate=True,
)
# Extract the generated token's log probabilities
scores = generation_output.scores # Tuple of length 1
logits = scores[0] # Tensor of shape (batch_size, vocab_size)
log_probs = F.log_softmax(logits, dim=-1) # Apply log softmax to get log probabilities
2024-11-25 11:00:03 -08:00
2024-11-25 15:36:04 -08:00
# Get top 5 tokens and their log probabilities
topk_log_probs, topk_indices = torch.topk(log_probs[0], k=5)
topk_tokens = self.tokenizer.convert_ids_to_tokens(topk_indices.tolist())
2024-11-25 11:00:03 -08:00
2024-11-25 15:36:04 -08:00
top_logprobs_hf.append((topk_tokens, topk_log_probs.tolist()))
2024-11-25 11:00:03 -08:00
2024-11-25 15:36:04 -08:00
# Pick the top token
next_token_id = topk_indices[0].unsqueeze(0).unsqueeze(0) # Shape: (1, 1)
next_token_str = self.tokenizer.convert_ids_to_tokens([next_token_id.item()])[0]
generated_tokens.append(next_token_id.item())
# Append the next token to input_ids and update attention_mask
2025-01-29 15:30:39 -08:00
inputs["input_ids"] = torch.cat([inputs["input_ids"], next_token_id], dim=-1)
inputs["attention_mask"] = torch.cat([inputs["attention_mask"], torch.ones((1, 1), dtype=inputs["attention_mask"].dtype).to(self.device)], dim=-1)
2024-11-25 11:00:03 -08:00
2024-12-03 15:32:53 -08:00
print(self.tokenizer.decode(generated_tokens))
2024-11-25 15:36:04 -08:00
# Now take all the input ids and run them through sglang as a comparison
async with AsyncClient(timeout=600) as session:
query["temperature"] = 0.0
2025-01-29 15:30:39 -08:00
query["max_tokens"] = max_steps
2024-11-25 15:36:04 -08:00
query["logprobs"] = True
query["top_logprobs"] = 5
COMPLETION_URL = f"http://localhost:{30000}/v1/chat/completions"
response = await session.post(COMPLETION_URL, json=query)
2024-11-25 10:12:29 -08:00
2024-11-25 15:36:04 -08:00
response_data = response.json()
2024-11-25 10:12:29 -08:00
2024-11-25 15:36:04 -08:00
for step, lptok in enumerate(response_data["choices"][0]["logprobs"]["content"]):
2024-11-25 16:08:24 -08:00
print("\nTop 5 tokens and their log probabilities:")
2024-11-25 15:36:04 -08:00
(topk_tokens, topk_log_probs) = top_logprobs_hf[step]
for token, log_prob, lptokcur in zip(topk_tokens, topk_log_probs, lptok["top_logprobs"]):
2025-01-29 15:30:39 -08:00
print(
f"HF Token: {token} Log Prob: {log_prob:.2f} Prob {math.exp(log_prob)*100:.2f}% SGLANG Token {lptokcur['token']} Logprob {lptokcur['logprob']:.2f} Prob {math.exp(lptokcur['logprob'])*100:.2f}%"
)
2024-11-25 10:32:18 -08:00
2024-11-25 10:12:29 -08:00
async def asyncTearDown(self):
# Clean up the model and tokenizer
del self.model
del self.tokenizer
2024-11-25 15:36:04 -08:00
torch.cuda.empty_cache()
2025-01-29 15:30:39 -08:00
2025-02-14 20:42:19 +00:00
@pytest.mark.nonci
2024-12-04 13:20:10 -08:00
class RawSGLangTest(unittest.IsolatedAsyncioTestCase):
def setUp(self):
# Set up the Hugging Face model and tokenizer
2025-01-29 15:30:39 -08:00
model_cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "olmocr", "model")
2024-12-04 13:20:10 -08:00
download_directory([MODEL_FINETUNED_PATH], model_cache_dir)
# Check the rope config and make sure it's got the proper key
with open(os.path.join(model_cache_dir, "config.json"), "r") as cfin:
config_data = json.load(cfin)
if "rope_type" in config_data["rope_scaling"]:
del config_data["rope_scaling"]["rope_type"]
config_data["rope_scaling"]["type"] = "mrope"
with open(os.path.join(model_cache_dir, "config.json"), "w") as cfout:
json.dump(config_data, cfout)
self.model_cache_dir = model_cache_dir
self.tokenizer = AutoTokenizer.from_pretrained(model_cache_dir, trust_remote_code=True)
self.image_token_id = self.tokenizer.encode("<|image_pad|>")[0]
self.model = Qwen2VLForConditionalGeneration.from_pretrained(model_cache_dir, torch_dtype=torch.bfloat16, trust_remote_code=True).eval()
self.processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.model.to(self.device)
# Path to the test PDF
self.test_pdf_path = Path(os.path.join(os.path.dirname(__file__), "gnarly_pdfs", "ambiguous.pdf"))
self.maxDiff = None
async def test_vision_encoder(self):
query = await build_page_query(
2025-01-29 15:30:39 -08:00
str(self.test_pdf_path),
page=1,
target_longest_image_dim=1024,
target_anchor_text_len=6000,
)
2024-12-04 13:20:10 -08:00
messages = query["messages"]
2025-01-29 15:30:39 -08:00
2024-12-04 13:20:10 -08:00
# Apply chat template to get the text
2025-01-29 15:30:39 -08:00
text = self.processor.apply_chat_template(query["messages"], tokenize=False, add_generation_prompt=True)
2024-12-04 13:20:10 -08:00
image_url = query["messages"][0]["content"][1]["image_url"]["url"]
# Remove the "data:image/png;base64," prefix
base64_image = image_url.split(",")[1]
# Decode the base64 string into bytes
image_data = base64.b64decode(base64_image)
# Create a BytesIO object and load it into a PIL image
main_image = Image.open(BytesIO(image_data))
# Process inputs using processor
inputs = self.processor(
text=[text],
images=[main_image],
padding=True,
return_tensors="pt",
)
with torch.no_grad():
2024-12-04 13:27:51 -08:00
hf_output = self.model.visual(inputs["pixel_values"].to(self.device), grid_thw=inputs["image_grid_thw"].to(self.device))
2024-12-04 13:20:10 -08:00
2024-12-04 13:27:51 -08:00
print("HF", hf_output, hf_output.shape)
2024-12-04 13:20:10 -08:00
from sglang.srt.configs.model_config import ModelConfig
2025-01-29 15:25:10 -08:00
from sglang.srt.hf_transformers_utils import get_tokenizer
2024-12-04 13:20:10 -08:00
from sglang.srt.managers.schedule_batch import Req, ScheduleBatch
from sglang.srt.model_executor.forward_batch_info import ForwardBatch
from sglang.srt.model_executor.model_runner import ModelRunner
from sglang.srt.sampling.sampling_params import SamplingParams
2025-01-29 15:25:10 -08:00
from sglang.srt.server_args import PortArgs, ServerArgs
2024-12-04 13:20:10 -08:00
2025-01-29 15:30:39 -08:00
model_config = ModelConfig(self.model_cache_dir, model_override_args="{}")
2024-12-04 13:20:10 -08:00
server_args = ServerArgs(model_path=self.model_cache_dir)
# Initialize model runner
model_runner = ModelRunner(
model_config=model_config,
mem_fraction_static=0.8,
gpu_id=0,
tp_rank=0,
tp_size=1,
nccl_port=12435,
server_args=server_args,
)
print(model_runner)
2024-12-04 13:27:51 -08:00
with torch.no_grad():
sglang_output = model_runner.model.visual(inputs["pixel_values"].to(self.device), grid_thw=inputs["image_grid_thw"].to(self.device))
print("SGLANG", sglang_output, sglang_output.shape)
2024-12-04 13:20:10 -08:00
2024-12-04 13:27:51 -08:00
# Convert to float32 for numerical stability if needed
hf = hf_output.float()
sg = sglang_output.float()
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
# Basic shape and dtype comparison
print("\n=== Basic Properties ===")
print(f"Shapes match: {hf.shape == sg.shape}")
print(f"HF shape: {hf.shape}, SGLang shape: {sg.shape}")
print(f"HF dtype: {hf.dtype}, SGLang dtype: {sg.dtype}")
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
# Move tensors to CPU for numpy operations
hf_np = hf.cpu().numpy()
sg_np = sg.cpu().numpy()
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
# Statistical metrics
print("\n=== Statistical Metrics ===")
print(f"Mean absolute difference: {torch.mean(torch.abs(hf - sg)).item():.6f}")
print(f"Max absolute difference: {torch.max(torch.abs(hf - sg)).item():.6f}")
print(f"Mean squared error: {torch.mean((hf - sg) ** 2).item():.6f}")
print(f"Root mean squared error: {torch.sqrt(torch.mean((hf - sg) ** 2)).item():.6f}")
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
# Cosine similarity (across feature dimension)
cos_sim = F.cosine_similarity(hf, sg)
print(f"Mean cosine similarity: {torch.mean(cos_sim).item():.6f}")
print(f"Min cosine similarity: {torch.min(cos_sim).item():.6f}")
# Find largest absolute differences
print("\n=== Largest Absolute Differences ===")
diffs = torch.abs(hf - sg)
flat_diffs = diffs.flatten()
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
# Get indices of top 10 differences
top_k = 10
top_values, top_flat_indices = torch.topk(flat_diffs, top_k)
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
# Convert flat indices to multidimensional indices
top_indices = np.unravel_index(top_flat_indices.cpu().numpy(), diffs.shape)
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
print(f"\nTop {top_k} largest absolute differences:")
print("Index".ljust(30) + "Difference".ljust(15) + "HF Value".ljust(15) + "SGLang Value")
print("-" * 75)
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
for i in range(top_k):
# Get the index tuple for this difference
idx = tuple(dim[i] for dim in top_indices)
diff_val = top_values[i].item()
hf_val = hf[idx].item()
sg_val = sg[idx].item()
2025-01-29 15:30:39 -08:00
2024-12-04 13:27:51 -08:00
# Format the index tuple and values
idx_str = str(idx)
2025-01-29 15:30:39 -08:00
print(f"{idx_str:<30}{diff_val:<15.6f}{hf_val:<15.6f}{sg_val:.6f}")