The new generation of Audi Q5 has appeared in China, with three screens, and is expected to be launched in China next year

Recently, a new generation has appeared in China, and the new long-axis version is expected to be launched in China next year, with power expected to be equipped with a 2.0T engine.

According to the spy photos and the overseas version, the appearance has become a lot more radical. The air intake grille is designed with a grid, and the outside is surrounded by chrome-plated trims. The new headlight shape is very sharp. The diversion grooves at both ends of the front surround are made of grid elements, making it very sporty.

On the side of the car body, the lines are very simple, the outside of the window is surrounded by chrome trim, and the double five-wheel hub adopts a color matching design.

In terms of tail shape, a through-type taillight design was replaced, and blackened mesh elements were added to the back.

In terms of interior, it has also been greatly upgraded, with a curved integrated screen design composed of 11.9-inch LCD instrument screen + 14.5-inch central control large screen, and a 10.9-inch passenger entertainment screen is also added. In terms of practical configuration, the new car has functions such as HUD head-up display, wireless charging of mobile phones, and adjustment of rear seat backrests.

In terms of power, the overseas version is equipped with 2.0T power and a 48V light mixing system, with a maximum power of 204 horsepower and 272 horsepower. The transmission is matched with a 7-speed wet dual-clutch transmission.

In the domestic medium-sized SUV market, the Audi Q5 has been performing very well, with average monthly sales exceeding 10,000 units. The new generation of Audi Q5 is more avant-garde and dynamic, with a significant improvement in interior technology, catering to the needs of the younger market. Therefore, after the domestic launch, continued hot sales should not be a big problem.

2024 "Her Economy" Insight Report

Lead:

QuestMobile data shows that the proportion of women in the mobile Internet has continuously increased to 49.5% in the past year. By January 2024, the monthly living scale of female users reached 608 million, up 1.9% year-on-year, among which, the post-90s generation accounted for the highest proportion, reaching 19.4%, and users in second-tier cities and above increased by 6.6% year-on-year.

Influenced by the structural changes of female users, the segmentation of female groups is more clear. The number of users such as urban white-collar workers, young treasure mothers, middle-class people in small towns and urban silver-haired women has reached 42.9 million, 71.72 million, 53.35 million and 65.71 million respectively. The consumption preferences of different groups are quite different. For example, young treasure mothers tend to be "exquisite parenting" and "face value consumption", and tend to improve with smart home products.

In fact, on the whole, the consumption power of female groups continues to improve. As of January 2024, the proportion of female users with online consumption power of more than 2,000 yuan reached 32.8%, an increase of 2.2% year-on-year; Influenced by the demand of fashion design and photo experience, OPPO, Huawei, Apple and vivo account for 23.5%, 21.4%, 20% and 16.3% respectively in the distribution of active terminals.

In addition, in terms of usage habits, sports and fitness, health management, photography, film and television dramas, etc. are all very sticky. Specifically, in terms of fields, mobile shopping and digital reading are the most popular applications, except that the average monthly time spent in social fields reaches 51.9 hours. In January 2024, the average monthly time spent by female users in these two fields reached 13.9 hours and 13.2 hours respectively, up by 38.5% and 11.3% respectively.

How to play specifically? Might as well read the report.

First, the overall online consumption capacity of female users has improved, while the offline consumption demand for local life and film performances has also maintained growth.

1. The proportion of female groups in the mobile Internet continues to rise, and the number of active users has increased to nearly 610 million.

2. The post-90s generation accounts for the highest proportion of the overall female internet users. In terms of inter-city distribution, the sinking market of the third-tier and below is still the main force, but it has shown a trend of shifting to high-tier cities.

3. The consumption power of women has increased year by year, and the high consumption power of more than 2,000 yuan has increased significantly.

QuestMobile data shows that in January 2024, the proportion of female users with online consumption capacity of more than 2,000 yuan was 32.8%, up 2.2% year-on-year.

4. In terms of purchase choice, female consumers pay more attention to fashion design, photo experience and entertainment functions, with OPPO becoming their first choice, and vivo and glory are also favored by female users.

5. Female users pay attention to sports, fitness and health management, like to take photos to record their lives at any time, have a wide interest in film and television dramas, and are keen on discussing and sharing while taking photos and chasing dramas, which shows strong social attributes.

6. The duration of female users’ contact with the Internet continues to increase, but the growth rate is lower than that of the whole network, and the number of apps used per capita is basically stable.

7. The stickiness of women’s use in many industries has increased, and the overall per capita use time of mobile shopping and mobile social APP has increased by more than 3 hours.

According to QuestMobile data, in January 2024, the average monthly usage time of mobile shopping APP for female users was as long as 13.9 hours, an increase of 38.5% compared with 10.1 hours in the same period last year.

8. The number of women in the digital reading industry has increased significantly, and their activity in social platforms has been further enhanced; At the same time of online activity, offline consumer demand such as local life and film performances has also maintained growth.

Second, the four female value groups show different consumption views.

1. Through various factors such as portraits, occupations and consumption characteristics, the huge female group is subdivided into multiple value groups.

2. Urban white-collar women: they prefer instant retail and member store mode, which is the main force of luxury consumption.

2.1 Urban white-collar women pay attention to the quality of personal life while working hard, and often use the fresh e-commerce platform. Traveling and vacationing is also a way for them to enjoy their lives.

2.2 Instant retail platforms such as Ding-dong grocery shopping and JD.COM’s arrival at home meet the immediate needs of women in the fast-paced life of the workplace, and urban white-collar women also love the member store model of Box Horse and Sam.

2.3 Urban white-collar women show strong consumption enthusiasm in the luxury goods market for the needs of the workplace and the concept of enjoying consumption.

2.4 In addition to standard coffee at work, functional tea and chain tea brands are rapidly becoming the new favorite drinks of female white-collar workers.

3, young treasure mom: high spending power, exquisite parenting, but also do not forget to meet their own "face value consumption" needs.

3.1 As a new generation of parenting groups, young mothers use the mother-infant platform to learn parenting knowledge, find scientific parenting methods, and pay attention to children’s preschool education and interest cultivation.

3.2 Young Bao Ma pursues exquisite child-rearing, and has more detailed and diverse needs in maternal and child consumption, and is less sensitive to the price of child-rearing products.

3.3 Self-satisfaction consumption and family consumption go hand in hand, and young mothers’ demand for "face value consumption" remains the same. While paying attention to maternal and child parent-child content, their enthusiasm for beauty cosmetics and fashion information has not diminished.

3.4 Faced with the dual pressures of child-rearing and housework, young treasure mothers tend to use advanced smart home products to improve housework efficiency and improve their quality of life.

4. Middle-class women in small towns: consumers pay more attention to cost performance and pursue personalization, and the demand for social entertainment is high.

4.1 Women in small towns love taking photos and retouching, and they are active in e-commerce shopping platforms, especially in discount e-commerce apps.

4.2 Sharing life through social media has become a normal state, and the use time of middle-class women in small towns in Tik Tok, Aauto Quicker and Xiaohongshu platforms has increased.

4.3 Compared with metropolises, small-town life often has less pressure on life and more leisure time, and has a strong willingness to spend on movies, fitness and tourism.

4.4 In terms of household appliances consumption, middle-class female users in small towns show a trend of intelligent and personalized consumption, and pay more attention to small household appliances such as beauty instruments and smart cleaning products such as sweeping robots.

5. Urban silver-haired women: developing in the direction of "digital life experts"

5.1 The degree of Internet use among urban silver-haired women has further deepened, and the duration and frequency of Internet use have increased.

5.2 Urban silver-haired women gradually develop online use habits, and the scale of APP users such as travel, entertainment, shopping and social interaction has increased significantly.

5.3 Urban silver-haired women are enthusiastic about surfing the Internet, and brushing videos and reading information are the main forms of entertainment in daily life.

5.4 Silver-haired women’s consumer goods are rich and diverse, and their family is still their consumption focus. With the change of consumption concept, silver-haired people have changed from just-needed consumption to quality life consumption.

Third, women’s consumption is more diversified, and marketing for "female power" has become a new trend.

1. "Other consumption" is gradually moving towards "TA consumption" without sexism. Taking automobile consumption as an example, with the rising demand of women to buy cars, more and more car companies are launching models featuring women.

In terms of mobile games, the new work "Love and Deep Space" of stacked paper games and "Beyond the World" of Netease games were launched at the same time, which started the dispute for female users.

2. The film and television works with the theme of "She" have sprung up. Recently, the drama series "The News Queen" and the film YOLO, which interpret women’s perspectives, have both broken the circle and set off a discussion upsurge on "She Power" on social media platforms.

3. The focus of brand marketing on Women’s Day on March 8th focuses on the export of values on topics such as feminism and women’s power, so as to show a richer and more diverse female image.

Topic: Asking for M8 booking orders to break through the 100,000-question boundary and redefine the new luxury standard with the first service.

  Since it was officially launched in 2021, it has won the trust of over 600,000 users with its solid product strength. In 2024, M7 and M9 in the market segment continued to sell well. In March this year, the customer service M8 officially opened the reservation. By April 6th, the customer service M8 had ordered more than 100,000 units. The hot-selling myth in the world of inquiry is still being written, and "wholeheartedly"

What should I do if sogou input method does not display words when typing?

What should I do if sogou input method does not display words when typing?

  Sogou input method is the input method that many friends like to use now, but sometimes, when using it, no words appear even though pinyin is typed and spaces are pressed, and many friends don’t know how to solve it. Let’s take a look at the solution that sogou input method does not display words when typing.

What should I do if sogou input method does not display words when typing?

Method 1:

  Sogou input method does not display the word selection box when typing. Please check whether there is any problem with the input method settings. Right-click on sogou Skin, set the properties, and see if 【 Hide status bar when full-screen program 】 is checked. Try removing it.

  In addition, in the settings shown below, check [Print by Space] to try.

Print by space

Method 2:

  The problem that sogou input method doesn’t display the word selection box is mostly found in various games, while sogou input method is prone to this problem. I suggest you try QQ input method, Baidu input method, or uninstall sogou input method and reinstall it.

Sogou input method does not display characters when typing.

  The above is all about what to do if sogou input method can’t see words. I hope it will help everyone.

Brief Introduction of sogou Input Method

  Sogou input method is a Chinese Pinyin input method recently launched by Sohu, which is a great breakthrough compared with Intelligent ABC and Microsoft Pinyin input method. First of all, it has increased a lot in the thesaurus. The sogou input method includes the most popular phrases, including the names of famous stars, into the thesaurus by analyzing a large number of web pages, so that the phenomenon that commonly used phrases can’t be typed is no longer seen. Secondly, the interface has also been greatly improved. Chinese characters will be automatically displayed below the typed letters, and it is no longer necessary to press the space bar. More convenient input and faster typing speed are the main features of sogou input method.

Sogou Pinyin Method

  Sogou input method is different from the traditional method in that it adopts search engine technology, which is the second generation input method. Thanks to the adoption of search engine technology, the speed has made a qualitative leap, and sogou is far ahead of other input methods in terms of the breadth of thesaurus and the accuracy of words.

  The biggest feature of sogou input method is the combination of input method and Internet. The input method will automatically update its own popular thesaurus, which is derived from the popular keywords of sogou search engine. In this way, the workload of users’ self-created words is reduced and the efficiency is improved.

Toyota Flanders/Nissan Hackers: Who is worth the money? be clear/manifest at a glance

I have seen a detailed comparison between Toyota and Nissan from the outside to the inside. So what about the engines of these two cars? What’s the difference between CVT gearboxes? Today we will open it for everyone to see!

engine

Flanders is equipped with a 2.0L self-priming four-cylinder engine, which is a classic M20 series with a maximum power of 126kW and a maximum torque of only 205nm.

Nissan hacker is equipped with a 1.3T turbocharged four-cylinder engine, model HR13, with a maximum power of 116kW and a maximum torque of 270Nm. This engine was actually developed by Nissan Renault in cooperation with Mercedes-Benz, so it can be seen in Class B and Nissan Renault cars.

This M20 engine is made of aluminum alloy block from top to bottom, and adopts Toyota D-4 mixed injection technology, which can be compatible with manifold injection and direct injection in cylinder at the same time, and the thermal efficiency of the whole engine is as high as 40%.

This M20 engine adopts laser side skirt piston, which improves the oil retention and scratch resistance. In addition, the cylinder is made of diamond-like carbon electroplating material and coated by physical vapor deposition technology to achieve the effects of high hardness, high elasticity, high inertia and low friction.

Hacker’s engine is also an all-aluminum cylinder from top to bottom. What is different is that this engine adopts a triangular cylinder head, which makes the exhaust gas closer to the engine center. The compact design makes the engine lose about 3 kilograms. In addition, the cylinder bore coating technology of Nissan’s new generation Nanoslide is adopted in the cylinder body, and the coating materials such as iron or carbon are instantly melted by high-voltage arc and evenly sprayed on the inner wall of lightweight aluminum alloy cylinder, so that the mechanical friction loss of the cylinder is reduced by 50% compared with that of the traditional gray cast iron cylinder liner, and at the same time, it has very high wear resistance.

In addition, the engine of hacker also uses an electronically controlled wastegate valve, which can change the opening pressure of the bypass valve. The later the opening time of the solenoid valve, the greater the boost pressure and the higher the power torque of the engine. In fact, this is how many engines adjust their high and low power. On the engine external characteristic curve, the speed at which the torque curve begins to decrease is the time for the bypass valve to open.

Looking at the exhaust part, Flanders uses a common exhaust manifold. Hacker is the latest integrated exhaust manifold. When the car is cold, the high temperature of the exhaust manifold can quickly heat the coolant, and when the engine is under high load, the coolant can also cool the exhaust. When the exhaust temperature is lowered, the pollutant content in the exhaust gas will be significantly reduced, which will help the engine meet more stringent environmental regulations. At the same time, because the integrated exhaust and cylinder body are almost integrated, the weight of the engine can be further reduced by about 3 kg.

The particle catcher in Flanders is far away from the engine, and the temperature is relatively low, which is not conducive to burning particles and is easy to block the particle catcher. The hacker’s particle catcher is close to the exhaust end of the engine, which can reduce the particle catcher by using the high temperature of the exhaust, making it difficult to block.

continuously variable transmission

Both Flanders and hackers match CVT gearboxes, but after disassembling, they are not the same.

Flanders use steel belt structure, and most CVT gearboxes are like this at present. The supplier of this steel belt is Bosch, which consists of numerous wedge-shaped steel sheets connected in series through two flexible steel belts. The biggest advantage of steel belt drive is low cost, and it can also reduce the size of gearbox. However, the disadvantage is that the maximum torque of power output will not be very large, and "slipping" may occur during rapid acceleration, and the instantaneous power transmission efficiency will become low. Moreover, it is sensitive to high temperature and low temperature, and the probability of steel strip fracture is relatively high.

Although hacker is also a CVT gearbox, it is no longer the previous Nissan CVT, and the steel belt drive is replaced by the steel chain drive of Schaeffler, Germany. Steel chain is countless perforated steel sheets connected by chain, and its strength and durability are much stronger than that of steel belt. The friction with the bevel wheel will be greater, and it can bear a large torque of 450 Nm, and the power loss during rapid acceleration is less, and it is not easy to "slip". However, the steel chain has high cost and relatively large volume, but it is more "anti-manufacturing", so it often appears in some cars that pay attention to power. For example, Audi A4 and Subaru XV all use CVT gearboxes driven by Schaeffler steel chain.

The CVT gearbox of Flanders uses Toyota’s "Drirect Shift-CVT" technology, that is, a gear is added in front of the CVT transmission wheel. When starting, the gear can be directly driven to avoid the loss of power due to the "slipping" of the starting steel belt, which is also a remedy for the weakness of its CVT gearbox.

Because the CVT of Fenglanda is driven by steel belt, it can’t output large torque, so only a mechanical oil pump is equipped to spray and lubricate the interior of the gearbox. The hacker uses steel chain transmission, which needs better heat dissipation, so the electronic high-pressure oil pump is used to provide higher oil pressure.

summary

Buy who? be clear/manifest at a glance

Through the dismantling of Toyota Fenglanda and Nissan hacker, Toyota is still the Toyota that pursues "enough is good". On the other hand, Nissan hackers are striving for change, "leaving Asia and entering Europe", using a higher-level Loulan platform, using a large number of aluminum alloys to lose weight, using a brand-new 1.3T turbocharged four-cylinder engine and using a steel chain CVT gearbox. To tell the truth, it’s really unexpected that Nissan can make such a change!

Then, how do you choose these two compact joint venture Japanese SUVs with similar prices? If you just want to buy a cheap, home, not too much pursuit of driving and motivation, then Feng Landa can do it. And if you want a young and fashionable appearance and high configuration, you have a certain pursuit of power and driving control. That’s obviously more value for money!(The conclusion of the measurement data in this article and the video is only responsible for this car, and the measurement data is responsible for the publicized experimental method. )

Total score of 2023 Toyota Flanders

Total score of 2023 Nissan hackers

Omicron is "sad" for spreading Christmas in Europe and America.

On December 24th, 2021, in new york, USA, passengers were waiting at the New York Airport. Hundreds of Christmas flights in the United States were cancelled due to the surge in confirmed cases in COVID-19 caused by the spread of Omicron strain.

  Affected by the rapid spread of Covid-19 Omicron strain, many airlines around the world canceled more than 4,500 flights during the Christmas holiday.

  European and American countries ushered in the Christmas holiday in the shadow of the epidemic. Recently, the number of confirmed cases in COVID-19 in the United States surged, and the number of newly confirmed cases in Britain and France reached a new high on the 24th.

  Thousands of flights were cancelled.

  Passenger transport market "adds congestion"

  "It’s so disappointing. After working hard for half a day, I still have to stay at home … …” Connie, a resident of Chicago, USA, told reporters in dismay after her flight was cancelled on the 24th. In order to fly to Atlanta to visit her parents during the Christmas holiday, Connie received a Covid-19 test and bought masks and other epidemic prevention supplies in advance. However, just before boarding the plane, she received the news that the flight was cancelled.

  Connie’s experience is not a case. Due to the rapid spread of Covid-19 Omicron strain in Europe and America, many staff members of passenger transport enterprises can’t continue to work because of diagnosis or close contact. European and American airline operators were forced to cancel thousands of flights, and many European railway operators also faced the same problem. For most people in Europe and America, Christmas is the most important holiday to reunite with their families in a year, but this year it is impossible to travel or the journey is interrupted due to the epidemic.

  According to the data of American flight tracking website, airlines around the world canceled at least 2,401 flights on the 24th, and nearly 10,000 flights were delayed. On Christmas Day, 25th, 1,779 flights worldwide were cancelled, and another 402 flights scheduled for 26th were cancelled.

  According to the above website data, among all cancelled flights, American international and domestic commercial flights account for more than a quarter. Among them, United Airlines and Delta Air Lines cancelled nearly 280 flights on the 24th, because of the shortage of manpower caused by the surge of infection cases in COVID-19.

  "The number of cases of Omicron virus infection in the United States has surged this week, which has had a direct impact on our pilots and operators," United said in a statement. "Therefore, we have to cancel some flights with regret."

  Delta Air Lines said that the company has tried every means to maintain the original flight, including changing the route and replacing the aircraft and crew. "We apologize for the delay in our customers’ holiday travel plans."

  Lufthansa said on the 24th that it cancelled more than a dozen long-distance transatlantic flights during the Christmas holiday due to the "substantial increase" in sick leave of pilots.

  Jetstar said that due to the spread of Covid-19, many staff members had to be tested and quarantined, which led to the last-minute delay or cancellation of flights. This enterprise has rebooked tickets for most passengers.

  According to Reuters, in Britain, many industries and transportation networks are facing a shortage of manpower because sick employees are isolating themselves.

  The epidemic in the United States has worsened.

  Shorten the isolation period of infected medical staff before returning to work

  Due to the spread of Omicron strain, the number of confirmed cases of COVID-19 in the United States has surged recently. According to a report by Reuters on the 24th, the daily average number of newly confirmed cases in the United States increased by 45% over the previous week, reaching 179,000. New York reported more than 44,000 new confirmed cases on the 24th, setting a new record. In addition, at least 10 states recorded a record number of new cases in a single day on the 23rd or 24th.

  According to the data of Johns Hopkins University, there are more than 51.96 million confirmed cases and 810,000 deaths in COVID-19.

  According to Reuters data, at present, Omicron cases account for about 75% of the newly confirmed cases of COVID-19 in the United States, and in some areas, the proportion reaches 90%.

  The increasing number of inpatients in COVID-19 has impacted the medical system in the Midwest of the United States, and the intensive care units in Indiana, Ohio and Michigan are preparing for the worst.

  In response to the worsening epidemic situation, the New Year celebration in Times Square, new york, USA, will be greatly scaled down on the 31st, and only a maximum of 15,000 people will be allowed to enter, so as to keep a safe distance.

  The U.S. Centers for Disease Control and Prevention recently issued a proposal that the isolation time of medical staff who tested positive in Covid-19 will be shortened before returning to work, in view of the shortage of manpower in hospitals across the United States.

  According to the latest recommendations of the CDC, medical staff infected with Covid-19 can return to work after 5 days of isolation if they have no symptoms or symptoms are mild and continue to improve, without providing negative COVID-19 test certificates; If the symptoms are serious after infection, you can’t return to work until 20 days after the symptoms appear. According to the CDC, if the shortage of medical staff increases, the isolation period may be further shortened.

  The CDC previously suggested that infected medical staff should be quarantined for 10 days from the day after symptoms or positive test results.

  Britain and France are hitting record highs.

  France may introduce new restrictions

  According to data released by the British government on the 24th, there were 122,186 new confirmed cases of COVID-19 in the UK in the past 24 hours, setting a new record since the outbreak, with a total of more than 11.89 million confirmed cases. There were 137 new deaths in a single day, with a cumulative death of more than 140,000.

  As of the week of 24th, there were about 700,000 confirmed cases and 810 deaths in Britain, increasing by 48.2% and 2% respectively compared with the previous week.

  According to the latest research report of Covid-19 released by the British Health and Safety Agency on 23rd, compared with the Delta strain, the Omicron strain spreads faster, but the risk of hospitalization of confirmed cases is lower. However, British health officials are cautious about the prospect of the epidemic.

  According to the data model of the Bureau of Statistics, the number of infected people in London will rise further in the next few days, and it is expected to increase to about one in every 10 people by the 26th.

  According to the data of the British Health and Safety Agency on 24th, there were 23,719 new cases of Omicron virus infection in the country, and 114,625 cases were confirmed. British Health Secretary Said Dzhavid pointed out earlier that there was a "data gap" in Omicron strain, and the actual number of infected cases was "significantly higher" than the official data.

  Dzhavid said on the same day that there are still many uncertainties surrounding the Omicron strain. Although the number of confirmed cases is increasing, it will not announce the tightening of prevention and control measures for the English epidemic after Christmas.

  The number of confirmed cases in COVID-19 has also reached a new high. On the 24th, 94,124 new cases were reported, with a total of 8.98 million confirmed cases. There were 167 new deaths, with a cumulative death of over 120,000. In the past seven days, 66,417 new cases were diagnosed daily in France, setting a new record, which tripled in one month.

  According to Reuters, Omicron virus infection cases have now accounted for 20% of newly confirmed cases in France.

  The French Presidential Palace said on the 24th that President Emmanuel Macron is scheduled to hold a special meeting on the COVID-19 epidemic on the 27th, and new epidemic prevention restrictions may be introduced.

  dynamic

  Tokyo is now the first

  Community-transmitted cases in Omicron

  Tokyo, Japan, 24 new cases of Covid-19 Omicron strain infection, including a 50-year-old male infected person who has neither left Japan recently nor been in close contact with immigrants, is regarded as the first case of Omicron community transmission in Tokyo.

  According to the report of Japan Broadcasting Association TV, this man is a doctor in a clinic in Tokyo and has been vaccinated with two doses of COVID-19 vaccine. He developed fever on 16th and was tested for Covid-19 nucleic acid, and the result was positive. The man was subsequently hospitalized, and there are no symptoms at present. Five close contacts, including his family and colleagues in the clinic, received nucleic acid tests, and the results are negative now.

  According to the Tokyo Metropolitan Government, men wear masks and protective masks when they see a doctor, so patients are not close contacts. However, considering the strong infectivity of Omicron strain, the Tokyo Metropolitan Government suggested that about 20 staff and 80 patients in the clinic should be screened for COVID-19.

  South Korea demands more than 200 people.

  "vacate" intensive care unit

  South Korea’s Central Anti-epidemic Countermeasures Headquarters reported on the 24th that the number of severe cases in COVID-19 nationwide reached 1,084, a new high since the outbreak. Due to the shortage of hospital beds, the Korean government ordered more than 200 patients who had been treated in the intensive care unit in COVID-19 for more than 20 days to leave, or they would be fined.

  The epidemic situation of COVID-19 in South Korea is tightening, and the number of newly confirmed cases in a single day in recent days is at a high level of around 7,000. According to the report of the Central Epidemic Prevention Countermeasures Headquarters on the 24th, there were 6,233 newly confirmed cases of COVID-19 in the past 24 hours, and 596,209 cases have been confirmed so far. There were 56 new deaths and 5071 cumulative deaths.

  With the number of severe cases in Korea hitting record highs, hospitals are overwhelmed and it is hard to find a bed. As of the early morning of 23rd, the utilization rate of intensive care beds in COVID-19 is about 80%, and that in the capital circle is as high as 85%.

  Faced with the shortage of beds, the Korean Ministry of Health and Welfare issued an administrative order on the afternoon of the 22nd, requiring 210 patients who have been treated in the intensive care unit for more than 20 days to leave, or they will be fined 1 million won (about 5,366 yuan).

  According to the regulations of the Ministry of Health and Welfare, if the patient is treated in the intensive care unit for more than 20 days, the subsequent treatment will no longer be free. At the same time, the relevant subsidies of the hospital where the patient is located will be cancelled.

  The Ministry of Health and Welfare said on the 23rd that as of that morning, 71 patients had left the intensive care unit and transferred to the general ward or other wards providing COVID-19 treatment.

  According to the data of the Central Epidemic Prevention Countermeasures Headquarters, as of 24th, 82.3% of the population in South Korea had been vaccinated with COVID-19 vaccine, and 27.9% had been vaccinated with booster shots.

  At the same time, the department warned that the mutant Covid-19 Omicron strain is spreading rapidly in South Korea, or will replace another variant Delta as the main epidemic strain in the next month or two. This group/Xinhua News Agency

The "room ticket compensation" for demolition funds reappears in the rivers and lakes, and many places have released big moves for the property market to go to inventory.

  In order to help urbanization, Zhengzhou is starting a new round of shed reform, and the latest housing ticket placement policy has attracted great attention.

  Recently, the General Office of Zhengzhou Municipal People’s Government issued the "Implementation Measures for Housing Ticket Resettlement in the Renovation Project of Big Shantytowns in Zhengzhou City (Provisional)", proposing that the renovation project of big shanty towns should implement housing ticket resettlement. The scope of application is the central city, that is, Zhongyuan District, Erqi District, Jinshui District, Guancheng Huizu District, Huiji District, Zhengdong New District, Zhengzhou Economic Development Zone and Zhengzhou High-tech Zone.

  The so-called house ticket, that is, after the monetary quantification of the expropriated person’s housing resettlement compensation rights and interests, the expropriated person issues a settlement certificate for the expropriated person to purchase a house. Previously, some cities in Henan and other provinces have implemented this resettlement method.

  Housing ticket placement, that is, in accordance with the policy of monetary placement and property right exchange, the collector converts the planned housing into money, and then issues it to the expropriated person in the form of housing ticket, and the latter buys commercial housing from the housing enterprises involved in the resettlement work.

  According to Zhengzhou policy, if the expropriated person chooses the house ticket for resettlement, it will be regarded as giving up the exchange of house property rights, and the expropriated person will be rewarded with 8% of the resettlement compensation rights and interests, and at the same time, the expropriated person will be rewarded with a three-month transition fee in cash.

  After receiving the house ticket, the range of commercial housing purchased by the expropriated person is not limited to the administrative area where the house ticket issuing institution is located, but can be selected by itself from the commercial housing provided by the housing enterprises participating in the house ticket resettlement work, and it is not included in the number of family purchase restrictions.

  In the purchase process, if the face value of the room ticket is enough to pay for the house purchase, it shall not be split and used. If it is not enough to pay the purchase price, the expropriated person may apply for commercial personal housing loans and provident fund personal housing loans.

  According to the document, the validity period of the house ticket is one year, and the expropriated person should pay for the house within the validity period. Those who are not used within the time limit shall be deemed to have given up the house ticket placement and re-arranged according to the monetary compensation method, but they will no longer enjoy the corresponding policy incentives for the house ticket placement.

  In addition, the house ticket is subject to real-name registration system, and the holder is the owner of the expropriated house, and the user is limited to the holder, the holder’s spouse and the holder’s close relatives. The room ticket shall not be transferred, donated, mortgaged, pledged or cashed out.

  Earlier on March 1, Zhengzhou had clearly stated in the Notice on Promoting the Benign Cycle and Healthy Development of the Real Estate Industry that it was necessary to implement the three-year action of resettlement housing construction, and adhere to the monetization resettlement based on the real estate market situation. For resettlement houses that have not started construction, the demolition people are encouraged to choose monetized resettlement.

  "The housing ticket placement policy itself is no stranger. It has been launched everywhere before, but in the current market environment, Zhengzhou’s housing ticket policy has new meaning." Some analysts in the industry said that buyers can buy houses through it, and the government still has certain rewards, which is equivalent to indirectly enhancing purchasing power. Buyers may buy larger units and high-quality houses, which is actually a destocking policy.

  The above analysts also said that the housing ticket policy has another meaning, which is actually a financial means. If you want to help inventory, just give money directly. Now it is equivalent to reducing direct compensation, and hope that funds will continue to flow into real estate. The government can not only control the cost of compensation and resettlement, but also encourage the expropriated households to purchase commercial housing by receiving "house tickets".

  According to public information, the cities that currently implement the housing ticket placement policy include Xinyang, Henan, Changshu, Jiangsu, Shaoxing, Zhejiang and Fenghua, Ningbo. From the time point of view, as early as 2015, this policy was introduced in various places. Since then, it has been used as a policy toolbox from time to time in various places, and the detailed rules have been continuously improved.

  Not long ago, Xinyang, Henan Province issued a new policy of "compensation for house expropriation and housing ticket placement", choosing the housing ticket compensation system, enjoying deed tax reduction, provident fund withdrawal and portfolio loan; Enjoy the compulsory education guarantee policy, you can retain the admission qualification of the original area, and you can also enjoy the admission qualification of the newly purchased commercial housing area; The room ticket is valid for 18 months.

  Fenghua District of Ningbo also started the room ticket policy again after 2016, after a lapse of six years, to promote the transformation process of the old city. The use of room tickets is limited to the whole area of Fenghua, and commercial houses and related ancillary facilities can be purchased, excluding stock houses (second-hand houses). The room ticket takes effect from the date of issuance and the redemption period is 12 months.

  At the moment of high real estate inventory, restarting the shed-to-house ticket placement system in many places means that a new round of destocking is quietly opening.

  Take Zhengzhou as an example. Since the beginning of this year, Zhengzhou has introduced the property market bail-out policy for several times in succession. Although the market activity has improved, the confidence of buyers, market vitality, and the de-commercialization of commercial housing need to be revitalized.

  According to the latest data from the National Bureau of Statistics, in April 2022, the price of new houses in Zhengzhou fell by 0.3% from the previous month, and the price of second-hand houses fell by 0.5% from the previous month. This is the ninth time that Zhengzhou’s housing market fell after the price of new houses fell in August last year.

  In terms of transactions, according to the data of Zhuge Housing Search Data Research Center, 6,892 sets of new commercial housing transactions were completed in Zhengzhou in April, down 11.25% from March; In May, 4,884 sets were sold, down 29.14% from last month and 73.54% year-on-year.

  On June 15th, Zhengzhou issued a document saying that commercial banks in Zhengzhou should be urged to implement the lower limit standard of 4.25% for the first home loan interest rate, and the down payment ratio for the second home should be reduced to 40%. The level of the above interest rates only appeared before 2016. Now, the housing ticket placement policy is offered, Zhengzhou toolbox is opened again, and the "destocking" of the property market is expected to accelerate.

Tongren Jietu X95 Price Reduction News! The highest discount is 0.5 million, and the discount waits for no one.

[car home Tongren Preferential Promotion Channel] Good news! This high-profile medium and large SUV is now launching unprecedented preferential activities, and consumers in tongren city can enjoy a car purchase discount of up to RMB 5,000. The original price of Jietu X95 has been adjusted to 141,900 yuan, which is very cost-effective. If you are interested in this luxurious and comfortable SUV, seize this rare opportunity and click "Check the car price" in the quotation form. Let’s explore more preferential possibilities together to help you realize your dream of buying a car.

铜仁捷途X95降价消息!最高优惠0.5万,优惠不等人

Jietu X95 leads the fashion trend with its magnificent design. In the front part, the family-style exquisite grille design is adopted, which gives the vehicle a sense of strength and recognition, and the streamlined outline lines show the perfect combination of sports and luxury. Overall style, it combines the elements of hard-core cross-country and urban business, creating a visual experience with a stable atmosphere and a sense of science and technology. Whether it is the details of the front or the proportional layout of the car body, it reflects the designer’s pursuit of quality and taste.

With its slender body size of 4858mm x 1925mm x 1780mm and wheelbase of 2850mm, Jetway X95 shows a steady and spacious body proportion. The lines on the side of the car are smooth and powerful, which outlines a unique dynamic aesthetic feeling. The front and rear wheel tracks are 1610mm and 1615mm respectively, which ensures the driving stability. Tyre size is 235/55 R19, which is consistent. With fashionable rim design, it adds a sense of movement and strength to the body.

铜仁捷途X95降价消息!最高优惠0.5万,优惠不等人

The interior design of Jietu X95 shows the perfect combination of exquisiteness and technology. The leather steering wheel not only provides a comfortable grip, but also supports manual up and down+front and rear adjustment to meet the needs of different drivers. The center console is embedded with a 12.3-inch large screen, which integrates the functions of multimedia system, navigation, telephone and air conditioning control, and cooperates with the voice recognition system, making the operation convenient and user-friendly. The seat part is made of imitation leather, and the main and passenger seats are equipped with front and rear adjustment, backrest adjustment and height adjustment to ensure the ride comfort. The second row of seats also supports front and rear and backrest adjustment, and has a proportional reclining function, providing greater flexibility for the seating space. These configurations fully reflect the practicality and luxury of Jietu X95 interior.

Jetway X95 is equipped with a 1.6T turbocharged engine with a maximum power of 145kW and a maximum torque of 290Nm. This powerful power source comes from the 197-horsepower L4 engine, and with the 7-speed wet dual-clutch gearbox, it provides a smooth power output and driving experience.

In the glorious chapter of Jetway X95, the constantly evolving and improving quality and cost performance make it maintain a strong competitiveness in the market. Now, in order to give back to the love of consumers, Jetway has announced the launch of a limited-time promotion. At this critical moment, you can not only have the excellent performance of this luxury SUV, but also enjoy unprecedented car purchase discounts. Seize this rare opportunity, let Jetway X95 become your effective partner in life, and let us witness this exciting moment together. Jietu looks forward to sailing with you to the future journey and opening a brand-new driving experience.

Will Tongyi divest Wu Jia from Alibaba Cloud or lead a team to integrate Ali AI To C business?

K graph BABA_0

K diagram 09988_0

  Recently, it was reported on the Internet that Ali’s AI application "Tongyi" was stripped from Alibaba Cloud. After adjustment, Tongyi App leveled the intelligent search product "Quark", while Tongyi Big Model will remain in Alibaba Cloud.

  Time Finance learned from many sources that Tongyi will be spun off from Alibaba Cloud. It is reported that Tongyi App is in the charge of Wu Jia, Quark and Tongyi App are two brands developed in parallel by the intelligent information enterprise group, and there is no change in ali tong Yiyi Lab and its Tongyi big model.

  It is understood that Wu Jia is currentlyVice President of the Group, President of Taotian User Platform, President of Ali Mama Business Department and President of Intelligent Information Business Group. After this year’s Tmall double 11 media communication meeting, Wu Jia also interviewed Time Finance and other media about the deployment of Taotian in double 11.

  On November 21 this year, Wu Yongming, CEO of Ali Group, issued an email to establish an e-commerce business group, with Jiang Fan as CEO. Since the end of November, there have been rumors in the market that the management team of Taobao Tmall has ushered in fine-tuning. There are rumors that Wu Jia has been transferred to Alibaba Cloud, and there are rumors that he has returned to Ali Group and will focus on exploring the business applied in AI to C. However, at present, Wu Jia’s next move has not yet been made public.

  According to public information, Wu Jia is one of the representatives of the "post-85 s" management of Ali Group. In 2010, Wu Jia joined Ali after graduation, during which she held several important positions.

  In 2017, Wu Jia became the general manager of UC Business Department of Ali Entertainment; In 2019, Ali Innovation Business Group was established, and Wu Jia was responsible for the business development and product innovation of UC Division and Shuqi Division. In 2021, Ali established an intelligent information business group on the basis of innovative business groups, with Wu Jia as the general manager, and in the same year, he was promoted to the president of Alibaba Intelligent Information Business Group. During the period, Wu Jia led the team to incubate C-end products that are popular among young people, such as Quark Browser and Quark Network Disk.

  At the end of last year, after a series of organizational restructuring in Alibaba, Wu Jia was appointed to be in charge of Taotian User Platform Division and Ali Mom Division.

  Quark is actually a major representative of Ali’s AI ToC business. In 2016, Quark was hatched by UC browser, focusing on minimalist, efficient and ad-free browsers. In 2023, Quark expanded many intelligent tools, including "Quark Listening", "Quark Scanning King" and "AI Natural Language Search". In January of this year, the official version of Quark PC browser was launched.

Source: Screenshot from Quark official website

  According to the statistics of the new list in November, Quark’s monthly visits exceeded 63 million. According to the latest data from PConline, Quark PC ranked first in the domestic computer AI application downloads in 2024. According to the report data released by CCID Research Institute, in the college entrance examination season in 2024, the use of quark AI search exceeded 100 million times, and the user scale reached a record high.

  Tongyi App is an all-round AI assistant developed by Alibaba Cloud, which integrates many advanced AI technologies such as natural language processing, speech recognition and image generation. According to QuestMobile’s data, as of July, the monthly user scale of AI native application (APP) has exceeded 66.3 million, among which Tongyi ranks in the top five, with a monthly user scale of 4.24 million.

Source: Screenshot from Tongyi Web Edition

  According to sources, after that, Alibaba Cloud focused more on To B business, and Tongyi App was classified as To C application in intelligent information business group. Although the latter was merged into the group, the underlying AI basic capability was still provided and supported by Alibaba Cloud.

  This year is the first year of AI application, and many large model manufacturers are burying their heads in building AI applications. Tongyi APP has also created some AI application functions that are highly adaptable to young people. For example, the "National Stage" channel launched by Tongyi App in April allows photos to act and sing; In August, Tongyi Web Edition launched the "PPT Creation" function; In November, Tongyi launched the online code mode, and an application can be generated in a vernacular, including small, data charts, websites, resumes, etc.

  This also echoes the organizational structure change of Alibaba Group last year. Since Wu Yongming took over as CEO of Ali Group in September last year, he has made clear the strategic focus and priority choice of the whole group, defined the core business and business strategy, and made firm choices around the main business. In November 2023, Ali announced the first batch of strategic innovative businesses-1688, idle fish, nail and quark-for young people and technology development trends, and made great efforts to promote the reform of young management team.

  In the future, with the upgrade of Tongyi App to Quark-level business, Ali Intelligent Information Business Group may become an important export of Ali AI To C applications.

AIGC’s landing threshold plummeted: the hardware budget was reduced to 1/46 in one breath, and one line of code was automatically parallelized.

Mingmin originated from Aofei Temple.Quantum bit | WeChat official account QbitAI
From AI painting to NLP big model, the landing cost of AIGC has been laid down at one time!
Not much to say, just look at the results:
Stable Diffusion 2.0 training/fine-tuning/reasoning can save up to 5.6 times of memory consumption, and make the hardware cost drop directly.1/46, a line of code can be enabled;BLOOM, a large model with 175 billion parameters, makes single-machine reasoning, which saves 4 times the memory consumption and reduces the hardware cost.One-tenth.One line of code realizes automatic search for the best parallel strategy, significantly reducing the threshold for distributed training, and natively supporting popular AI model libraries such as Hugging Face and Timm.
You know, on the other side of the AIGC explosion, the high cost plagues the whole industry.
Last week, StockAI, one of the first AI painting companies, was forced to announce the closure of the platform. For no other reason, the founder said:
The driving cost of the company is too high, and the current income is unsustainable.
Even with the support of OpenAI and Microsoft behind ChatGPT, an announcement was issued a few weeks after the platform was opened, and users began to limit the daily use times.
The implication is nothing more than four words:Can’t afford to burn.
In a word, reducing the landing cost of AI big model is an urgent problem in the industry at present.
At the same time, the open source AI big model solutionColossal-AIIn the past year, it has quickly become popular, and 7k+Star has been harvested on GitHub.
If you drop this plan as above, it will come from its hands.
How exactly is it achieved? Look down,
Open source address:https://github.com/hpcaitech/ColossalAI
Stable Diffusion2.0 low-cost training/fine-tuning/reasoning
Compared with version 1.0, Stable Diffusion 2.0 not only improves the pixels of the image generated by the model, but also introduces Depth2img model, text-guided repair model and so on, with more perfect functions.
This wave of novelty actually surprised and caught users off guard.
(After all, I haven’t understood 1.0 yet)
But then again, it’s still an old problem. The cost of landing AIGC model is high.
Take Stable Diffusion as an example. The Stability AI behind it maintains more than 4,000 GPU clusters of NVIDIA A100, and has spent more than 50 million dollars in operating costs.
Faced with fast iterative models, algorithms and downstream tasks,How to reduce the application costIt has become the core issue of AIGC’s real landing.
Stable Diffusion 2.0 is built on the easy-to-use PyTorch Lightning framework.
As the official big model solution of PyTorch Lightning, Colossal-AI followed up for the first time.
The specific contents are as follows:
The memory consumption can be saved by 5.6 times, and the hardware cost can be reduced to 1/46 at most.Support DreamBooth single GPU for fast personalized fine tuning.Reasoning memory consumption saves 2.5 times.Moreover, the scheme will also be merged into Hugging Face in the near future, which will further facilitate users’ use.
train
In order to speed up the training and reduce the training cost,Use a larger batch sizeIt has become an effective means widely used. However, the limited memory capacity of GPU seriously limits the size of batch size and pushes up the threshold of training hardware.
Through a series of video memory optimization techniques, Colossal-AI enables Stable Diffusion to use the video memory requirements of large batch size 16 training on each GPU on average.From 64.5GB to 11.6GB., saving 5.6 times, and can also be extended to single GPU or multi-GPU parallel.
Compared with using the most advanced A100 80GB, it only needs3060 and other consumer graphics cardsCan meet the demand,The hardware cost has dropped to 1/46 at the highest..
Therefore, more users can carry out the research and application of Stable Diffusion at low cost on consumer GPU.
Back-memory optimization
Flash Attention
As early as version 1.0 of Stable Diffusion, Colossal-AI took the lead in introducing Flash Attention technology and succeeded.Increase the speed of attention by 104%, will be end-to-end trainingPeak memory reduced by 23%.
Flash Attention is an accelerated version of long sequence attention, which uses Flat to reduce the number of memory reads/writes between high-bandwidth memory (HBM) of GPU. At the same time, Flash attention designs an approximate attention algorithm for massive sparse attention, which is faster than any existing approximate attention method.
In version 1.0 of stable diffusion, there are only a few attention layers in the whole Diffusion Model, and Flash attention has not shown its performance advantages.
In Stable Diffusion 2.0, a large number of convolution layers are replaced by attention layers, which further exerts the memory optimization potential of Flash Attention.
ZeRO + Gemini
Colossal-AI supports the method of ZeRO redundancy optimizer to eliminate memory redundancy. Compared with the classical data parallelism strategy, Colossal-AI can greatly improve the efficiency of memory use without sacrificing computational granularity and communication efficiency.
In addition, Colossal-AI also introducedChunk mechanismFurther improve the performance of ZeRO.
A continuous set of parameters in the operation order is stored in a Chunk (Chunk is a continuous memory space), and each Chunk has the same size.
Chunk organizing memory can ensure the efficient use of network bandwidth between PCI-e and GPU-GPU, reduce communication times and avoid potential memory fragmentation.
However, Colossal-AI’s heterogeneous memory space manager Gemini supports offloading optimizer state from GPU to CPU to save GPU memory occupation.
GPU memory and CPU memory (composed of CPU DRAM or NVMe SSD memory) can be used at the same time to break through the limitation of single GPU memory wall and further expand the scale of trainable model.
One line of code to get started quickly
As an official partner of PyTorch Lightning,Only one line of code is needed to call the above video memory optimization of Colossal-AI..
from lightning.pytorch import trainer, LightningModulefrom lightning.pytorch.strategies import ColossalAIStrategyMystrategy = ColossalAIStrategy(use_chunk=True, enable_distributed_storage=True, placement_policy=auto)trainer = Trainer(accelerator="gpu", devices=4, precision=16, strategy=Mystrategy)trainer.fit(model)
DreamBooth fine-tuning
Colossal-AI also released the fine-tuning scheme of DreamBooth model "conveniently" while launching the acceleration scheme of Stable Diffusion 2.0.
This is the model released by Google in August this year. It just needs to3-5Pictures, together with words, can make the specified objects migrate to other scenes or styles.
The biggest difference from Dall-E 2, Imagen, etc. is that DreamBooth can select objects.Faithful reduction.
In the scheme, users only need to run the file train_dreambooth_colossalai.py directly, so that they can give full play to the memory optimization of Colossal-AI in this fine-tuning task, and fine-tune their graphic model personalized and quickly, which greatly reduces the use threshold.
reason
Because model reasoning is insensitive to numerical accuracy, it is possible to realize low-precision and low-cost reasoning.
For the Stable Diffusion 2.0 model, you can use theAdd a line of code, support the model’s Int8 quantitative reasoning,Memory consumption is reduced by 2.5 times, and only 3.1GB is needed.And doe not cause significant performance loss.
model = replace_module(model)
Reasoning 175 billion BLOOM model with RTX3090
On the other side of AI painting explosion, the trend of NLP big model continues.
In July this year, Hugging Face released an open source model with 175 billion parameters.BLOOMIt took 384 pieces of A100 to make it.
If the common FP32/FP16 is directly used for reasoning, and 8 GPUs in a single node use model parallelism,Each GPU needs to consume at least 87.5GB/43.8GB of video memory..
Such a large memory occupation, even the most advanced 8-card A100(80GB/40GB) server, alsoInference service cannot be deployed directly.And multi-node reasoning will bring heavy extra cost and communication overhead.
Based on this situation, Colossal-AI realizes efficient Int8 quantization and model parallel reasoning, which can serve the reasoning of large models such as BLOOM with 175 billion parameters,8-card server deployed to consumer graphics cards such as 3090/4090.At the same time, it does not produce significant CPU memory occupation improvement and performance loss.
Compared with the original A100 scheme, the hardware deployment cost can be reduced to the original one.One-tenth.
By quantizing the model with Int8, Colossal-AI can reduce the total memory occupation of the model from 352.3GB(FP16) to 185.6GB. At the same time, using the model parallel technology of Colossal-AI, it willThe occupation of each graphics card is reduced to 23.2GB..
In model parallelism, in order not to increase the CPU memory occupation, Colossal-AI quantifies and segments the model in the main process, and lazy_init is used in each other to obtain the meta model which occupies little memory and video memory, and then the model parameters are passed between processes through gloo backend.
Through the above scheme, the CPU memory occupation can reach the peak without loading the model parameters in stages, reaching the theoretical optimal level. Compared with the "pipeline-like" distribution mode in which the model is divided into layers, model parallelism can improve the efficiency of memory utilization under non-intensive requests.
One line of code automatically parallelizes.
The distributed mixed deployment of large models is a very complicated problem.
At present, the common distributed large-scale model training schemes all rely on repeated attempts by users and configuration and deployment by system experts based on experience.
However, this is very unfriendly to most AI developers, because everyone doesn’t want to spend too much time and energy on studying distributed systems and trial and error.
Thus, Colossal-AI’s efficient and easy-to-use automatic parallel system can be said to solve everyone’s urgent needs.
Just add one line of code., it can provide cluster information and single-machine training model to obtain distributed training ability, andNative support includes popular AI model libraries such as Hugging Face and Timm..
# wrap the model using auto_enginemodel, optimizer = auto_engine(model, optimizer, cluster_info)# normal training loop
Therefore, Colossal-AI can greatly reduce the threshold for AI developers to use distributed technology to train and fine-tune large models. At the same time, the automatic parallel system can find a more efficient parallel scheme from the finer-grained search parallel strategy.
Graph Tracing
Colossal-AI isThe first automatic parallel system based on PyTorch framework using static graph analysis.
PyTorch is a dynamic graph framework, and obtaining its static execution plan is a long-term research problem in the field of machine learning system.
Colossal-AI uses ColoTracer based on torch.FX Tracer to derive and record the meta-information of each tensor during tracing, such as tensor shape, dims, dtype, etc., which can help the subsequent automatic parallel strategy search.
Therefore, Colossal-AI has better model generalization ability, instead of relying on model name or manual modification to adapt parallel strategy.
Fine-grained distributed training strategy search
Colossal-AI will search the strategy for each op with the fastest running time as the goal under the limitation of memory budget, and finally get the real training strategy, including the segmentation strategy of each tensor, the types of communication operators to be inserted between different computing nodes, whether to replace operators, etc.
Tensor parallelism and data parallelism in the existing system, and mixed parallelism such as column segmentation and row segmentation used by NVIDIA in parallel systems such as Megatron-LM are all subsets of the strategies that can be searched by automatic parallelism.
In addition to these parallel modes that can be specified manually, Colossal-AI’s automatic parallel system has the ability to specify a unique parallel mode for each op, so it is possible to find a better parallel strategy than manual segmentation that relies on expert experience and trial-and-error configuration.
Distributed tensor and shape consistency System
Similar to the latest release of DTensor by PyTorch, Colossal-AI also uses device mesh to manage the cluster abstractly.
Specifically, Colossal-AI uses sharding spec to label the distributed storage state of tensor, and uses shape consistency manager to automatically convert the same tensor between different sharding specs.
This greatly improves the versatility and ease of use of Colossal-AI. With the help of shape consistency manager, tensor can be segmented without burden, without worrying that the output of upstream op and the input of downstream OP are stored differently in the cluster.
Compared with PyTorch DTensor, Colossal-AI has the following three advantages:
The device mesh of Colossal-AI can profiling the cluster performance index and estimate the time consumption of different communication operators.The shape consistency of Colossal-AI will greedily search for the conversion mode between Harding Specs, instead of simply converting one dimension at a time, so as to find a more efficient conversion path, and thus make the conversion communication overhead between Harding Specs lower.Adding all_to_all operation makes Colossal-AI more extensible, which can show great advantages when training on large-scale clusters.Combined with activation checkpoint.
As an essential memory compression technology in large model training, Colossal-AI also provides an automatic search function for activation checkpoint.
Compared with most technical solutions that aim at maximum video memory compression, Colossal-AI’s search goal is to find the fastest activation checkpoint solution within the video memory budget.
At the same time, in order to avoid the explosion of search time caused by modeling the search of activation checkpoint into SPMD solver, Colossal-AI has designed 2-stage search, so it can search for an effective and feasible distributed training scheme in a reasonable time.
About Colossal-AI
Colossal-AI, a universal deep learning system, faces the era of big model, which can realize efficient and rapid deployment of AI big model training and reasoning, and reduce the application cost of AI big model.
Since the open source, Colossal-AI has been ranked first in the GitHub hot list for many times, with over 7,000 GitHub Star, and has been successfully selected as the official course of international AI and HPC top conferences such as SC, AAAI and PPoPP.
Colossal-AI related solutions have been successfully applied by well-known manufacturers in autonomous driving, cloud computing, retail, medicine, chips and other industries, and received wide acclaim.
For example, the recent explosion of fireChatGPT, not yet open source and does not have networking function. Colossal-AI has successfully helped a Fortune 500 company to develop a chat bot model with enhanced online search engine capabilities.
portal
Open source address:https://github.com/hpcaitech/ColossalAI
Reference link:https://www.hpc-ai.tech/blog/colossal-ai-0-2-0
Reporting/feedback