[DO NOT MERGE] Upstream test PR #322
Closed
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
Removes unnecessary mark step from MoE OP loop to speed up computation
Signed-off-by: Chendi.Xue <chendi.xue@intel.com>
Signed-off-by: Chendi.Xue <chendi.xue@intel.com>
Work around for allocation error while loading llama-405b.
This bugfix addresses incorrect lower boundary handling for bucketing Previous behavior: ``` INFO 09-03 19:36:28 habana_model_runner.py:564] Prompt bucket config (min, step, max_warmup) bs:[64, 32, 64], seq:[768, 128, 768] INFO 09-03 19:36:28 habana_model_runner.py:577] Generated 12 prompt buckets: [(32, 128), (32, 256), (32, 384), (32, 512), (32, 640), (32, 768), (64, 128), (64, 256), (64, 384), (64, 512), (64, 640), (64, 768)] INFO 09-03 19:36:28 habana_model_runner.py:582] Omitted 0 prompt buckets due to exceeded token budget (max_num_batched_tokens=131072) INFO 09-03 19:36:28 habana_model_runner.py:590] Decode bucket config (min, step, max_warmup) bs:[64, 128, 64], seq:[768, 128, 1024] INFO 09-03 19:36:28 habana_model_runner.py:601] Generated 8 decode buckets: [(64, 128), (64, 256), (64, 384), (64, 512), (64, 640), (64, 768), (64, 896), (64, 1024)] INFO 09-03 19:36:28 habana_model_runner.py:606] Omitted 0 decode buckets due to exceeded token budget (max_num_batched_tokens=131072) ``` Min seq len dimension is set to 768, but buckets with seq_len=128-768 are present Current behavior: ``` INFO 09-03 19:45:42 habana_model_runner.py:563] Prompt bucket config (min, step, max_warmup) bs:[64, 32, 64], seq:[768, 128, 768] INFO 09-03 19:45:42 habana_model_runner.py:576] Generated 1 prompt buckets: [(64, 768)] INFO 09-03 19:45:42 habana_model_runner.py:581] Omitted 0 prompt buckets due to exceeded token budget (max_num_batched_tokens=131072) INFO 09-03 19:45:42 habana_model_runner.py:589] Decode bucket config (min, step, max_warmup) bs:[64, 128, 64], seq:[768, 128, 1024] INFO 09-03 19:45:42 habana_model_runner.py:600] Generated 3 decode buckets: [(64, 768), (64, 896), (64, 1024)] INFO 09-03 19:45:42 habana_model_runner.py:605] Omitted 0 decode buckets due to exceeded token budget (max_num_batched_tokens=131072) ``` No bucket with seq_len < 768 is captured
Signed-off-by: Chendi.Xue <chendi.xue@intel.com>
This PR prevents max_num_batched_tokens from limiting decode buckets, as decode buckets should be limited by number of blocks, not by max_num_batched_tokens.
Ports #97 to habana_main
Refactors BGMV implementation from gather based to mask-based to optimize performance and reduce device memory usage.
Use all possible slot values for dummy blocks to avoid caching issues.
With PT_COMPILE_ONLY_MODE flag, graphs can be compiled without performing synLaunch. The flag has been added to the warmup phase to decrease its execution time.
This fixes a very silly issue where mismatching values of `warmup_mode` flag could cause graph recompilations and eventually memory leaks.
This PR fixes crashes observed on older Synapse builds introduced with #227. Setting PT_COMPILE_ONLY_MODE is not supported in current or older public Synapse builds, but we should not crash because of it, rather we should advise user to use the latest build. Previous behavior: ``` ... INFO 09-06 17:08:37 habana_executor.py:85] # HPU blocks: 10761, # CPU blocks: 910 INFO 09-06 17:08:37 habana_worker.py:201] Initializing cache engine took 47.29 GiB of device memory (54.34 GiB/94.62 GiB used) and -159.6 MiB of host memory (414.9 GiB/1007 GiB used) [rank0]: Traceback (most recent call last): [rank0]: File "/software/users/kzawora/vllm-utils/vllm_hpu_simple_test.py", line 9, in <module> [rank0]: llm = LLM(model="facebook/opt-125m") [rank0]: File "/software/users/kzawora/vllm-fork/vllm/entrypoints/llm.py", line 155, in __init__ [rank0]: self.llm_engine = LLMEngine.from_engine_args( [rank0]: File "/software/users/kzawora/vllm-fork/vllm/engine/llm_engine.py", line 456, in from_engine_args [rank0]: engine = cls( [rank0]: File "/software/users/kzawora/vllm-fork/vllm/engine/llm_engine.py", line 266, in __init__ [rank0]: self._initialize_kv_caches() [rank0]: File "/software/users/kzawora/vllm-fork/vllm/engine/llm_engine.py", line 378, in _initialize_kv_caches [rank0]: self.model_executor.initialize_cache(num_gpu_blocks, num_cpu_blocks) [rank0]: File "/software/users/kzawora/vllm-fork/vllm/executor/habana_executor.py", line 89, in initialize_cache [rank0]: self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks) [rank0]: File "/software/users/kzawora/vllm-fork/vllm/worker/habana_worker.py", line 202, in initialize_cache [rank0]: self._warm_up_model() [rank0]: File "/software/users/kzawora/vllm-fork/vllm/worker/habana_worker.py", line 220, in _warm_up_model [rank0]: self.model_runner.warmup_model(self.hpu_cache[0]) [rank0]: File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context [rank0]: return func(*args, **kwargs) [rank0]: File "/software/users/kzawora/vllm-fork/vllm/worker/habana_model_runner.py", line 1412, in warmup_model [rank0]: with compile_only_mode_context(): [rank0]: File "/usr/lib/python3.10/contextlib.py", line 135, in __enter__ [rank0]: return next(self.gen) [rank0]: File "/usr/local/lib/python3.10/dist-packages/habana_frameworks/torch/internal/bridge_config.py", line 20, in env_setting [rank0]: get_func = globals()['get_' + var.lower()] [rank0]: KeyError: 'get_pt_compile_only_mode' inc shutdown inc shutdown inc shutdown inc shutdown ``` Current behavior: ``` ... INFO 09-06 17:06:42 habana_executor.py:85] # HPU blocks: 10761, # CPU blocks: 910 INFO 09-06 17:06:43 habana_worker.py:201] Initializing cache engine took 47.29 GiB of device memory (54.34 GiB/94.62 GiB used) and -143.7 MiB of host memory (415 GiB/1007 GiB used) WARNING 09-06 17:06:43 habana_model_runner.py:1419] Cannot use PT_COMPILE_ONLY_MODE. Warmup time will be negatively impacted. Please update Gaudi Software Suite. INFO 09-06 17:06:43 habana_model_runner.py:1336] [Warmup][Prompt][1/23] batch_size:2 seq_len:1024 free_mem:40.28 GiB ... ```
Fixes serving mode issue; due to error in fastapi
This PR contains mask based BGMV implementation for LoRA embedding instead of index-select of LoRA-B weights. Removing special handling in no LoRA case also.
Eliminate two graph breaks for torch.compile mode:
1. [__graph_breaks] torch._dynamo.exc.Unsupported: builtin: eq [<class
'torch._dynamo.variables.misc.GetAttrVariable'>, <class
'torch._dynamo.variables.constant.EnumVariable'>] False
2. [__graph_breaks] torch._dynamo.exc.Unsupported: Tensor.item
---
<details>
<!-- inside this <details> section, markdown rendering does not work, so
we use raw html here. -->
<summary><b> PR Checklist (Click to Expand) </b></summary>
<p>Thank you for your contribution to vLLM! Before submitting the pull
request, please ensure the PR meets the following criteria. This helps
vLLM maintain the code quality and improve the efficiency of the review
process.</p>
<h3>PR Title and Classification</h3>
<p>Only specific types of PRs will be reviewed. The PR title is prefixed
appropriately to indicate the type of change. Please use one of the
following:</p>
<ul>
<li><code>[Bugfix]</code> for bug fixes.</li>
<li><code>[CI/Build]</code> for build or continuous integration
improvements.</li>
<li><code>[Doc]</code> for documentation fixes and improvements.</li>
<li><code>[Model]</code> for adding a new model or improving an existing
model. Model name should appear in the title.</li>
<li><code>[Frontend]</code> For changes on the vLLM frontend (e.g.,
OpenAI API server, <code>LLM</code> class, etc.) </li>
<li><code>[Kernel]</code> for changes affecting CUDA kernels or other
compute kernels.</li>
<li><code>[Core]</code> for changes in the core vLLM logic (e.g.,
<code>LLMEngine</code>, <code>AsyncLLMEngine</code>,
<code>Scheduler</code>, etc.)</li>
<li><code>[Hardware][Vendor]</code> for hardware-specific changes.
Vendor name should appear in the prefix (e.g.,
<code>[Hardware][AMD]</code>).</li>
<li><code>[Misc]</code> for PRs that do not fit the above categories.
Please use this sparingly.</li>
</ul>
<p><strong>Note:</strong> If the PR spans more than one category, please
include all relevant prefixes.</p>
<h3>Code Quality</h3>
<p>The PR need to meet the following code quality standards:</p>
<ul>
<li>We adhere to <a
href="https://google.github.io/styleguide/pyguide.html">Google Python
style guide</a> and <a
href="https://google.github.io/styleguide/cppguide.html">Google C++
style guide</a>.</li>
<li>Pass all linter checks. Please use <a
href="https://github.com/vllm-project/vllm/blob/main/format.sh"><code>format.sh</code></a>
to format your code.</li>
<li>The code need to be well-documented to ensure future contributors
can easily understand the code.</li>
<li>Include sufficient tests to ensure the project to stay correct and
robust. This includes both unit tests and integration tests.</li>
<li>Please add documentation to <code>docs/source/</code> if the PR
modifies the user-facing behaviors of vLLM. It helps vLLM user
understand and utilize the new features or changes.</li>
</ul>
<h3>Notes for Large Changes</h3>
<p>Please keep the changes as concise as possible. For major
architectural changes (>500 LOC excluding kernel/data/config/test), we
would expect a GitHub issue (RFC) discussing the technical design and
justification. Otherwise, we will tag it with <code>rfc-required</code>
and might not go through the PR.</p>
<h3>What to Expect for the Reviews</h3>
<p>The goal of the vLLM team is to be a <i>transparent reviewing
machine</i>. We would like to make the review process transparent and
efficient and make sure no contributor feel confused or frustrated.
However, the vLLM team is small, so we need to prioritize some PRs over
others. Here is what you can expect from the review process: </p>
<ul>
<li> After the PR is submitted, the PR will be assigned to a reviewer.
Every reviewer will pick up the PRs based on their expertise and
availability.</li>
<li> After the PR is assigned, the reviewer will provide status update
every 2-3 days. If the PR is not reviewed within 7 days, please feel
free to ping the reviewer or the vLLM team.</li>
<li> After the review, the reviewer will put an <code>
action-required</code> label on the PR if there are changes required.
The contributor should address the comments and ping the reviewer to
re-review the PR.</li>
<li> Please respond to all comments within a reasonable time frame. If a
comment isn't clear or you disagree with a suggestion, feel free to ask
for clarification or discuss the suggestion.
</li>
</ul>
<h3>Thank You</h3>
<p> Finally, thank you for taking the time to read these guidelines and
for your interest in contributing to vLLM. Your contributions make vLLM
a great tool for everyone! </p>
</details>
---------
Signed-off-by: yuwenzho <yuwen.zhou@intel.com>
FILL IN THE PR DESCRIPTION HERE
FIX #xxxx (*link existing issues this PR will resolve*)
**BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE
DESCRIPTION ABOVE**
---
<details>
<!-- inside this <details> section, markdown rendering does not work, so
we use raw html here. -->
<summary><b> PR Checklist (Click to Expand) </b></summary>
<p>Thank you for your contribution to vLLM! Before submitting the pull
request, please ensure the PR meets the following criteria. This helps
vLLM maintain the code quality and improve the efficiency of the review
process.</p>
<h3>PR Title and Classification</h3>
<p>Only specific types of PRs will be reviewed. The PR title is prefixed
appropriately to indicate the type of change. Please use one of the
following:</p>
<ul>
<li><code>[Bugfix]</code> for bug fixes.</li>
<li><code>[CI/Build]</code> for build or continuous integration
improvements.</li>
<li><code>[Doc]</code> for documentation fixes and improvements.</li>
<li><code>[Model]</code> for adding a new model or improving an existing
model. Model name should appear in the title.</li>
<li><code>[Frontend]</code> For changes on the vLLM frontend (e.g.,
OpenAI API server, <code>LLM</code> class, etc.) </li>
<li><code>[Kernel]</code> for changes affecting CUDA kernels or other
compute kernels.</li>
<li><code>[Core]</code> for changes in the core vLLM logic (e.g.,
<code>LLMEngine</code>, <code>AsyncLLMEngine</code>,
<code>Scheduler</code>, etc.)</li>
<li><code>[Hardware][Vendor]</code> for hardware-specific changes.
Vendor name should appear in the prefix (e.g.,
<code>[Hardware][AMD]</code>).</li>
<li><code>[Misc]</code> for PRs that do not fit the above categories.
Please use this sparingly.</li>
</ul>
<p><strong>Note:</strong> If the PR spans more than one category, please
include all relevant prefixes.</p>
<h3>Code Quality</h3>
<p>The PR need to meet the following code quality standards:</p>
<ul>
<li>We adhere to <a
href="https://google.github.io/styleguide/pyguide.html">Google Python
style guide</a> and <a
href="https://google.github.io/styleguide/cppguide.html">Google C++
style guide</a>.</li>
<li>Pass all linter checks. Please use <a
href="https://github.com/vllm-project/vllm/blob/main/format.sh"><code>format.sh</code></a>
to format your code.</li>
<li>The code need to be well-documented to ensure future contributors
can easily understand the code.</li>
<li>Include sufficient tests to ensure the project to stay correct and
robust. This includes both unit tests and integration tests.</li>
<li>Please add documentation to <code>docs/source/</code> if the PR
modifies the user-facing behaviors of vLLM. It helps vLLM user
understand and utilize the new features or changes.</li>
</ul>
<h3>Notes for Large Changes</h3>
<p>Please keep the changes as concise as possible. For major
architectural changes (>500 LOC excluding kernel/data/config/test), we
would expect a GitHub issue (RFC) discussing the technical design and
justification. Otherwise, we will tag it with <code>rfc-required</code>
and might not go through the PR.</p>
<h3>What to Expect for the Reviews</h3>
<p>The goal of the vLLM team is to be a <i>transparent reviewing
machine</i>. We would like to make the review process transparent and
efficient and make sure no contributor feel confused or frustrated.
However, the vLLM team is small, so we need to prioritize some PRs over
others. Here is what you can expect from the review process: </p>
<ul>
<li> After the PR is submitted, the PR will be assigned to a reviewer.
Every reviewer will pick up the PRs based on their expertise and
availability.</li>
<li> After the PR is assigned, the reviewer will provide status update
every 2-3 days. If the PR is not reviewed within 7 days, please feel
free to ping the reviewer or the vLLM team.</li>
<li> After the review, the reviewer will put an <code>
action-required</code> label on the PR if there are changes required.
The contributor should address the comments and ping the reviewer to
re-review the PR.</li>
<li> Please respond to all comments within a reasonable time frame. If a
comment isn't clear or you disagree with a suggestion, feel free to ask
for clarification or discuss the suggestion.
</li>
</ul>
<h3>Thank You</h3>
<p> Finally, thank you for taking the time to read these guidelines and
for your interest in contributing to vLLM. Your contributions make vLLM
a great tool for everyone! </p>
</details>
---------
Co-authored-by: Michal Adamczyk <madamczyk@habana.ai>
Co-authored-by: barak goldberg <149692267+bgoldberg-habana@users.noreply.github.com>
Co-authored-by: Michal Szutenberg <37601244+szutenberg@users.noreply.github.com>
Co-authored-by: Jan Kaniecki <jkaniecki@habana.ai>
RuntimeErrors are not observed anymore on habana_main when disable_tensor_cache is used. This PR enables disable_tensor_cache.
This PR raises the allowed relative tolerance in GSM8K to 0.06, and moves Llama-70B test to 4xG2 from 2xG2 until memory usage is investigated (success run: vLLM-CI-Pipeline/206)
We were asked on upstream PR to remove our changes from cache_engine.py. This PR does just that, and creates HPUCacheEngine inheriting from CacheEngine, just overriding _allocate_kv_cache method.
…imit prefill batch size (#394) This PR adds following functionality that can be enabled via engine flags: - use_padding_aware_scheduling - vLLM scheduler will now calculate token cost considering padded prefill shape (similar to #109). - max_num_prefill_seqs - padding-aware scheduler will perform an additional check for prefill batch size and will effectively limit prefill batch size at maximum of `max_num_prefill_seqs`. If unset, max prefill batch size will be `max_num_seqs`. Both features are generic and do not require HPU, although they may be specialized for particular vendor's usage. Padding aware scheduling includes padding function selector which selects HPU padding function (considering currently used HPU buckets) if current device is HPU. Otherwise, it will take a product of batch_size x max_seq_len.
|
Upstream finished, closing. |
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.
Changes w.r.t. habana_main:
mark_step()injections from models - we were asked to do so