[RFC/WIP] First steps towards FP8 for Mixtral #3208
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Instructions on how to generate activation scales with https://github.com/AniZpZ/AutoSmoothQuant: And then run (you need to download the pile validation data file |
|
I uploaded the model with quantized MoE layers to https://huggingface.co/pcmoritz/Mixtral-8x7B-Instruct-v0.1-FP8MOE now, it can be used with this PR. It needs to be loaded with I'd love to hear about people's experience with this if anybody tries it out :) |
pcmoritz
added a commit
that referenced
this pull request
Apr 24, 2024
This PR is the first step towards fixing #3208 It implements dynamic per-tensor scaling (see #4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this: ```python from vllm import LLM, SamplingParams prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8") outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` **Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in #3954). With this PR, the results are as follows: <img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03"> **Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows: ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7018|± |0.0036| | - humanities |N/A |none | 5|acc |0.6472|± |0.0065| | - other |N/A |none | 5|acc |0.7673|± |0.0072| | - social_sciences|N/A |none | 5|acc |0.8099|± |0.0070| | - stem |N/A |none | 5|acc |0.6131|± |0.0083| ``` this compares favorably with the fp16 results which are ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7020|± |0.1313| | - humanities |N/A |none | 5|acc |0.6425|± |0.1349| | - other |N/A |none | 5|acc |0.7744|± |0.1038| | - social_sciences|N/A |none | 5|acc |0.8131|± |0.0695| | - stem |N/A |none | 5|acc |0.6108|± |0.1383| ``` Happy hacking!
xjpang
pushed a commit
to xjpang/vllm
that referenced
this pull request
Apr 25, 2024
This PR is the first step towards fixing vllm-project#3208 It implements dynamic per-tensor scaling (see vllm-project#4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this: ```python from vllm import LLM, SamplingParams prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8") outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` **Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in vllm-project#3954). With this PR, the results are as follows: <img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03"> **Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows: ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7018|± |0.0036| | - humanities |N/A |none | 5|acc |0.6472|± |0.0065| | - other |N/A |none | 5|acc |0.7673|± |0.0072| | - social_sciences|N/A |none | 5|acc |0.8099|± |0.0070| | - stem |N/A |none | 5|acc |0.6131|± |0.0083| ``` this compares favorably with the fp16 results which are ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7020|± |0.1313| | - humanities |N/A |none | 5|acc |0.6425|± |0.1349| | - other |N/A |none | 5|acc |0.7744|± |0.1038| | - social_sciences|N/A |none | 5|acc |0.8131|± |0.0695| | - stem |N/A |none | 5|acc |0.6108|± |0.1383| ``` Happy hacking!
robertgshaw2-neuralmagic
pushed a commit
to neuralmagic/nm-vllm
that referenced
this pull request
Apr 26, 2024
This PR is the first step towards fixing vllm-project#3208 It implements dynamic per-tensor scaling (see vllm-project#4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this: ```python from vllm import LLM, SamplingParams prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8") outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` **Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in vllm-project#3954). With this PR, the results are as follows: <img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03"> **Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows: ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7018|± |0.0036| | - humanities |N/A |none | 5|acc |0.6472|± |0.0065| | - other |N/A |none | 5|acc |0.7673|± |0.0072| | - social_sciences|N/A |none | 5|acc |0.8099|± |0.0070| | - stem |N/A |none | 5|acc |0.6131|± |0.0083| ``` this compares favorably with the fp16 results which are ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7020|± |0.1313| | - humanities |N/A |none | 5|acc |0.6425|± |0.1349| | - other |N/A |none | 5|acc |0.7744|± |0.1038| | - social_sciences|N/A |none | 5|acc |0.8131|± |0.0695| | - stem |N/A |none | 5|acc |0.6108|± |0.1383| ``` Happy hacking!
alexeykondrat
pushed a commit
to alexeykondrat/ci-vllm
that referenced
this pull request
May 1, 2024
This PR is the first step towards fixing vllm-project#3208 It implements dynamic per-tensor scaling (see vllm-project#4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this: ```python from vllm import LLM, SamplingParams prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8") outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` **Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in vllm-project#3954). With this PR, the results are as follows: <img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03"> **Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows: ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7018|± |0.0036| | - humanities |N/A |none | 5|acc |0.6472|± |0.0065| | - other |N/A |none | 5|acc |0.7673|± |0.0072| | - social_sciences|N/A |none | 5|acc |0.8099|± |0.0070| | - stem |N/A |none | 5|acc |0.6131|± |0.0083| ``` this compares favorably with the fp16 results which are ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7020|± |0.1313| | - humanities |N/A |none | 5|acc |0.6425|± |0.1349| | - other |N/A |none | 5|acc |0.7744|± |0.1038| | - social_sciences|N/A |none | 5|acc |0.8131|± |0.0695| | - stem |N/A |none | 5|acc |0.6108|± |0.1383| ``` Happy hacking!
joerunde
pushed a commit
to IBM/vllm
that referenced
this pull request
May 6, 2024
This PR is the first step towards fixing vllm-project/vllm#3208 It implements dynamic per-tensor scaling (see vllm-project/vllm#4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this: ```python from vllm import LLM, SamplingParams prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8") outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` **Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in vllm-project/vllm#3954). With this PR, the results are as follows: <img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03"> **Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows: ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7018|± |0.0036| | - humanities |N/A |none | 5|acc |0.6472|± |0.0065| | - other |N/A |none | 5|acc |0.7673|± |0.0072| | - social_sciences|N/A |none | 5|acc |0.8099|± |0.0070| | - stem |N/A |none | 5|acc |0.6131|± |0.0083| ``` this compares favorably with the fp16 results which are ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7020|± |0.1313| | - humanities |N/A |none | 5|acc |0.6425|± |0.1349| | - other |N/A |none | 5|acc |0.7744|± |0.1038| | - social_sciences|N/A |none | 5|acc |0.8131|± |0.0695| | - stem |N/A |none | 5|acc |0.6108|± |0.1383| ``` Happy hacking!
shaojiewang
pushed a commit
to shaojiewang/vllm-rocm
that referenced
this pull request
Jul 3, 2024
This PR is the first step towards fixing vllm-project/vllm#3208 It implements dynamic per-tensor scaling (see vllm-project/vllm#4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this: ```python from vllm import LLM, SamplingParams prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8") outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` **Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in vllm-project/vllm#3954). With this PR, the results are as follows: <img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03"> **Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows: ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7018|± |0.0036| | - humanities |N/A |none | 5|acc |0.6472|± |0.0065| | - other |N/A |none | 5|acc |0.7673|± |0.0072| | - social_sciences|N/A |none | 5|acc |0.8099|± |0.0070| | - stem |N/A |none | 5|acc |0.6131|± |0.0083| ``` this compares favorably with the fp16 results which are ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7020|± |0.1313| | - humanities |N/A |none | 5|acc |0.6425|± |0.1349| | - other |N/A |none | 5|acc |0.7744|± |0.1038| | - social_sciences|N/A |none | 5|acc |0.8131|± |0.0695| | - stem |N/A |none | 5|acc |0.6108|± |0.1383| ``` Happy hacking!
Temirulan
pushed a commit
to Temirulan/vllm-whisper
that referenced
this pull request
Sep 6, 2024
This PR is the first step towards fixing vllm-project#3208 It implements dynamic per-tensor scaling (see vllm-project#4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this: ```python from vllm import LLM, SamplingParams prompts = [ "Hello, my name is", "The president of the United States is", "The capital of France is", "The future of AI is", ] sampling_params = SamplingParams(temperature=0.8, top_p=0.95) llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8") outputs = llm.generate(prompts, sampling_params) # Print the outputs. for output in outputs: prompt = output.prompt generated_text = output.outputs[0].text print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") ``` **Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in vllm-project#3954). With this PR, the results are as follows: <img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03"> **Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows: ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7018|± |0.0036| | - humanities |N/A |none | 5|acc |0.6472|± |0.0065| | - other |N/A |none | 5|acc |0.7673|± |0.0072| | - social_sciences|N/A |none | 5|acc |0.8099|± |0.0070| | - stem |N/A |none | 5|acc |0.6131|± |0.0083| ``` this compares favorably with the fp16 results which are ``` | Groups |Version|Filter|n-shot|Metric|Value | |Stderr| |------------------|-------|------|-----:|------|-----:|---|-----:| |mmlu |N/A |none | 0|acc |0.7020|± |0.1313| | - humanities |N/A |none | 5|acc |0.6425|± |0.1349| | - other |N/A |none | 5|acc |0.7744|± |0.1038| | - social_sciences|N/A |none | 5|acc |0.8131|± |0.0695| | - stem |N/A |none | 5|acc |0.6108|± |0.1383| ``` Happy hacking!
|
This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you! |
|
This pull request has merge conflicts that must be resolved before it can be |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Even after the optimizations #2979 and #2542, most of the time in the Mixtral forward pass is still spent in the fully connected layer / fused_moe kernel, see e.g. the following profiling run:
The next step to optimize this will be to use fp8 for these computations, which is known to perform well (https://arxiv.org/pdf/2208.09225.pdf, https://blog.fireworks.ai/fireattention-serving-open-source-models-4x-faster-than-vllm-by-quantizing-with-no-tradeoffs-a29a85ad28d0, https://arxiv.org/pdf/2303.17951.pdf).
NOTE: This is not yet working end-to-end, i.e. the numerics are not correct yet. We will most likely need to introduce scaling factors like in https://arxiv.org/abs/2211.10438
Can be tried out with
Current performance on master: 4.40s
Current performance on this PR: 3.40s
This PR has a long way to go before it can be ready. I'd love to get the communities input on what the best open source quantization codebases are that we could leverage to compute the scaling factors (e.g. I'm aware of https://github.com/AniZpZ/AutoSmoothQuant and https://github.com/mit-han-lab/smoothquant).