Removes unnecessary mark step from MoE OP loop to speed up computation

Signed-off-by: Chendi.Xue <[email protected]>

Signed-off-by: Chendi.Xue <[email protected]>

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 <[email protected]>

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 <[email protected]>

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 <[email protected]>
Co-authored-by: barak goldberg <[email protected]>
Co-authored-by: Michal Szutenberg <[email protected]>
Co-authored-by: Jan Kaniecki <[email protected]>

RuntimeErrors are not observed anymore on habana_main when
disable_tensor_cache is used. This PR enables disable_tensor_cache.

On habana_main the slots are calculated by adding an offset to the block
which breaks the check for _PAD_SLOT_ID. Reworked it so that in case of
_PAD_BLOCK_ID we're automatically inserting the right value.

Porting PT Profiler from:

81a23a7
and

e805b88

…a/vllm_v0_6_0_rebase

…m_v0_6_0_rebase

This PR handles mask creation inside lora unit tests to align with new
BGMV implementation

Attn MetaData was hard coded to bfloat16, leading to a runtime error for
float32 model instantiation.

Fixed ALiB and [MPT-7B](https://www.databricks.com/blog/mpt-7b) model.
Accuracy results comparing to CPU(collected using
[EleutherAI](https://github.com/EleutherAI/lm-evaluation-harness))

| Tasks          | CPU    | HPU    |
| -------------- | ------ | ------ |
| arc_challenge  | 0.4224 | 0.4189 |
| arc_easy       | 0.6974 | 0.6999 |
| hellaswag      | 0.7603 | 0.7626 |
| lambada_openai | 0.7306 | 0.7326 |
| mmlu           | 0.293  | 0.2925 |
| winogrande     | 0.6851 | 0.6811 |

Fixing ENV variables' names after flat-PA merge

This PR removes the hardcoded value used to normalize softmax in flat_pa
. Current approach is to use the global maximum as it is very easy to
compute, but it has the drawback that other samples in a batch might
slightly affect numerical stability.

This is a first step to eliminated some of the INF/NaN issues we see in
certain configurations and by no means this is a complete solutions.
This needs to be revised in the future.

Signed-off-by: Chendi.Xue <[email protected]>

Signed-off-by: Chendi.Xue <[email protected]>

Signed-off-by: Chendi.Xue <[email protected]>

Increase garbage collector's threshold in order to reduce it's frequency

FILL IN THE PR DESCRIPTION HERE

FIX ##198

After this change, we can see tool_calls can be returned successfully

``` bash
Compiling FSM index for all state transitions: 100%|████████████████████████████████████████████████████████████████████████| 55/55 [00:01<00:00, 32.86it/s]INFO 09-04 02:15:34 metrics.py:406] Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Swapped: 0 reqs, Pending: 0 reqs, GPU KV cache usage: 0.0%, CPU KV cache usage: 0.0%.
INFO 09-04 02:15:34 logger.py:36] Received request chat-0fd03b03ae05473488d9bce566401d91: prompt: "<|im_start|>user\nWhat's the weather like in Boston today?<|im_end|>\n<|im_start|>assistant\n", params: SamplingParams(n=1, best_of=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.0, temperature=0.7, top_p=1.0, top_k=-1, min_p=0.0, seed=None, use_beam_search=False, length_penalty=1.0, early_stopping=False, stop=[], stop_token_ids=[], include_stop_str_in_output=False, ignore_eos=False, max_tokens=1000, min_tokens=0, logprobs=None, prompt_logprobs=None, skip_special_tokens=True, spaces_between_special_tokens=True, truncate_prompt_tokens=None), prompt_token_ids: [27, 91, 318, 5011, 91, 29, 882, 198, 3923, 596, 279, 9282, 1093, 304, 10406, 3432, 76514, 91, 318, 6345, 91, 397, 27, 91, 318, 5011, 91, 29, 78191, 198], lora_request: None, prompt_adapter_request: None.
INFO 09-04 02:15:34 async_llm_engine.py:173] Added request chat-0fd03b03ae05473488d9bce566401d91.
INFO 09-04 02:15:36 async_llm_engine.py:140] Finished request chat-0fd03b03ae05473488d9bce566401d91.
INFO:     127.0.0.1:40452 - "POST /v1/chat/completions HTTP/1.1" 200 OK
Message: ChatCompletionMessage(content='', refusal=None, role='assistant', function_call=None, tool_calls=[ChatCompletionMessageToolCall(id='chatcmpl-tool-af3eac9372144f959ed0df7e16cf5da4', function=Function(arguments='{ "location": "Boston, MA", "unit": "fahrenheit" }', name='get_current_weather'), type='function')])
```

**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>

FILL IN THE PR DESCRIPTION HERE

for model(like chatglm2/3-6b) whose `rotary_dim` not equal to
`head_size`, current code will crash due to dim not equal.
#212 have a not robust enough fix. chatglm series could work, but
chatglm2-6b result is not correct.
this fix follow vllm rotary_embeding pytorch native impl. verified on
chatglm2-6b and chatglm3-6b

**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>

…ve CUDA hardcode (#217)

FILL IN THE PR DESCRIPTION HERE

FIX #216

**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>

Add Dockerfile.hpu

FIX #199

**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>

Signed-off-by: Chendi.Xue <[email protected]>

Signed-off-by: Chendi.Xue <[email protected]>

## One line description

Use topk instead of sort for topp/topk calculation under certain
conditions (scalar value of p and k).

## Details

Instead of using `k` for topk, we use `_padded_k`, which is strictly
larger than k and monotonically non decreasing.

We need/use `_padded_k > k` for cases where the smallest value of the
topk=k values has some values beyond k, (for example for
[9,8,8,8,7,7,7], with k=3, we have [9,8,8,8], which is 4 instead of 3
values),

To prevent excessive recompilations, anytime we require an expansion of
`_padded_k` we increment with a fixed constant `_increment` (usually
>1), to have a bucketed approach to prevent multiple shapes


### Basic outline

1. perform topk with `_padded_k`
2. find the "kth" value in each row (smallest number that will be in
topk), this is variable `num_duplicates_of_smallest_of_topk`
3. find maximum of number of duplicates, this variable is
`max_num_duplicates_of_smallest_of_topk`
4. check if `_padded_k` is big enough to contain
`max_num_duplicates_of_smallest_of_topk`. if not, then expand
`_padded_k`, and redo the topk again with expanded `_padded_k`
6. maskout the values that are extra in `_padded_k`
7. move to doing topp


## Perf benefit

### Using benchmark_throughput.py

To check benefit of this PR, make following change in
`benchmark_throughput.py`:
```
diff --git a/benchmarks/benchmark_throughput.py b/benchmarks/benchmark_throughput.py
index ff33e3dc..3383dea8 100644
--- a/benchmarks/benchmark_throughput.py
+++ b/benchmarks/benchmark_throughput.py
@@ -116,8 +116,9 @@ def run_vllm(
         sampling_params.append(
             SamplingParams(
                 n=n,
-                temperature=0.0 if use_beam_search else 1.0,
-                top_p=1.0,
+                temperature=1.0,  #0.0 if use_beam_search else 1.0,
+                top_p=0.95,
+                top_k=20,
                 use_beam_search=use_beam_search,
                 ignore_eos=True,
                 max_tokens=output_len,

 ```


`VLLM_SKIP_WARMUP=true VLLM_GRAPH_RESERVED_MEM=0.2 VLLM_GRAPH_PROMPT_RATIO=0.8 VLLM_DECODE_BS_BUCKET_MIN=1 VLLM_DECODE_BLOCK_BUCKET_STEP=64 VLLM_DECODE_BLOCK_BUCKET_MIN=64 python benchmark_throughput.py --model /root/sasarkar/llama3-8b/ --device hpu --seed 2024 --backend vllm --num-prompts 100 --dtype bfloat16 --input-len=256 --output-len=512`

in the numbers below there is a **49%** increase in thruput in the case with warmup, and **30%** increase in thruput in the case without warmup


#### with opt + warmup

Processed prompts: 100%|█████████████████████████████████████████████████████████████████████| 100/100 [00:22<00:00,  4.37it/s, est. speed input: 1119.66 toks/s, output: 2239.33 toks/s]
Throughput: 4.37 requests/s, 3354.58 tokens/s


#### with opt + skip warmup

Processed prompts: 100%|██████████████████████████████████████████████████████████████████████| 100/100 [00:46<00:00,  2.17it/s, est. speed input: 556.32 toks/s, output: 1112.63 toks/s]
Throughput: 2.17 requests/s, 1667.89 tokens/s


#### without opt + warmup

Processed prompts: 100%|██████████████████████████████████████████████████████████████████████| 100/100 [00:34<00:00,  2.93it/s, est. speed input: 749.24 toks/s, output: 1498.48 toks/s]
Throughput: 2.92 requests/s, 2245.74 tokens/s


#### without opt + skip warmup

Processed prompts: 100%|███████████████████████████████████████████████████████████████████████| 100/100 [00:59<00:00,  1.67it/s, est. speed input: 428.49 toks/s, output: 856.99 toks/s]
Throughput: 1.67 requests/s, 1284.85 tokens/s

### Using server Client
(Data collected by Peter)
[baseline](https://github.com/HabanaAI/vllm-fork/commits/a7763a7a76b4531ed7907549724df2949d9225bf/)
all numbers on 1.17-495
third column [branch ](https://github.com/HabanaAI/vllm-fork/commits/ae_benchmark_9_10_24/)

| model | TP | baseline HPU thruput   | baseline HPU + this PR thruput | baseline HPU + this PR + other opt | 
| -------- | ------- | ------- | ------- | ------- |
| llama3 8b | 1 | 950  | 1296    | 1306 | 
| llama3 8b | 4 | 1347  | 1969    | 2077 | 
| llama3 70b | 4 | 368  | 394    | 394 | 
| qwen 72b | 4 | 731  | 726    | 815 |


### Without delayed sampling 
On habana_main f858d43
```VLLM_GRAPH_RESERVED_MEM=0.2 VLLM_GRAPH_PROMPT_RATIO=0.8
VLLM_DECODE_BS_BUCKET_MIN=1 VLLM_DECODE_BLOCK_BUCKET_STEP=64
VLLM_DECODE_BLOCK_BUCKET_MIN=64 python benchmark_throughput.py --model
/root/sasarkar/llama3-8b/ --device hpu --seed 2024 --backend vllm
--num-prompts 100 --dtype bfloat16 --input-len=256 --output-len=512```

Without change
Throughput: 3.32 requests/s, 2550.85 tokens/s

With change:
Throughput: 5.17 requests/s, 3967.58 tokens/s




## Extra Notes
1. Works only for "scalar" case, though it might be possible to extend
the basic idea (topk instead of sort) for vector case as well. (Outline
of this is: find max k in topk vector, then perform topk using that,
etc. needs some bucketing possibly to prevent dyn shapes etc)
2. Need an additional check in `_init_sampling_tensors` to determine if
its scalar case. This has a minor perf hit. ideally if someone could
tell us that its a scalar from the top itself...
3. Some tradeoffs can be made, where we use a sufficiently large
padded_k (which is still smaller than vocab size) from the beginning,
and hope that every case lands within that bucket. Cases that wont land
are expected to be very, very rare. For example if padded_k = max(2 * k,
100) is used, and k = say 50, then we need the smallest of the topk
value to repeat 50 times with same probability, which is exceedingly
unlikely. If we trade off this mathematical improbability, then we can
do with only 1 topk op, which might be faster
4. There is a `fliplr` in the code, which could be removed, if we can
compute reverse cumsum. however the formula for reverse cumsum as
expressed [here ](pytorch/pytorch#33520), ` x
+ torch.sum(x, dim=1, keepdims=True) - torch.cumsum(x, dim=1)` is
numerically unstable, because of the addition/subtraction. It works well
enough on ints and large numbers, but not on small probability values.
5. The value of `k` affects the gains we might get from this. For
example in the expt shown above, with k=20, thruput increases from
1284.85 to 1667.89 (30% gain). But if k = 2000, instead of 20,
throughput increases from 1127.34 to 1289.26 (14% gain). Thus the gain %
might decrease with increasing k, as asymptotically topk would probably
converges to sort's performance for large k. However practically k is
pretty small.
6. For larger models, the gains may be less, as they are more device
bound probably
7. Cumsum may be taking long. Maybe try below. [Initial
try](b392ff8)
```
import torch
y = torch.tensor([[1,2,3], [4,5,6]])
mask1 = torch.tensor([[[1,0,0],[1,1,0],[1,1,1]], [[1,0,0],[1,1,0],[1,1,1]]])
torch.sum(y.unsqueeze(1)*mask1,2)
```
or
```
F.conv1d(torch.tensor([[[0,0,0,0,1,2,3,4,5]], [[0,0,0,0,6,7,8,9,10.0]]]), torch.ones([1,1,5], dtype=torch.float32))
```
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>

…lava-1.5 (#283)

FILL IN THE PR DESCRIPTION HERE

FIX #282  (*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>

…odule. (#250)

Co-authored-by: Konrad Zawora <[email protected]>

Update documentation on support of fp8

Moved files from vllm/hpu to another public repo:
https://github.com/HabanaAI/vllm-hpu-extension
It can be installed with pip install
git+https://github.com/HabanaAI/vllm-hpu-extension.git

After #252, HPUGraph capture
takes much less memory, and we can reduce the memory reserved for
HPUGraphs. On Llama3.1-8b-Instruct (G2), capturing 100% of prefill and
decode graphs on BS=256 now takes 1.566 GB of HBM, which is far less
than 40% (~30 GB) we reserve by default. This results in lots of unused
(==wasted) memory, which could be used instead for more KV cache blocks.

The original code doesn't print the default value correctly

INFO 09-17 00:06:07 habana_model_runner.py:95]
VLLM_PROMPT_BS_BUCKET_MIN=1 (default:_**min**_)
INFO 09-17 00:06:07 habana_model_runner.py:95]
VLLM_PROMPT_BS_BUCKET_STEP=1 (default:_**step**_)
INFO 09-17 00:06:07 habana_model_runner.py:95]
VLLM_PROMPT_BS_BUCKET_MAX=1 (default:_**max**_)

This change make it print the correct default value
INFO 09-17 21:30:51 habana_model_runner.py:95]
VLLM_PROMPT_BS_BUCKET_MIN=1 (default:_**1**_)
INFO 09-17 21:30:51 habana_model_runner.py:95]
VLLM_PROMPT_BS_BUCKET_STEP=4 (default:_**32**_)
INFO 09-17 21:30:51 habana_model_runner.py:95]
VLLM_PROMPT_BS_BUCKET_MAX=4 (default:_**64**_)

Fix blocks number calculation for Flat PA via adding empty table_block
(#158)

Remove dummy seq metadata from loop for Flat PA fix

Add extra mark_step() on each decode layer to optimize the performance
on Gaudi.

Signed-off-by: Bob Zhu <[email protected]>

…ng (#297)

fix bug: device_str in initialize_ray_cluster requires uppercase string

w/o the bug fix, multi HPUs will encounter "ValueError: The number of
required hpus exceeds the total number of available hpus in the
placement group" error, as the device_str is not expected as uppercase,
then available hpus always returns 0.

Fixes Lora Related issues in vllm Rebase

Recent changes broke slot sparsity for warmup slots. This commit
restores the functionality.

…313)

Block 0 is used for padding. This PR removes the padding block from a
list of available blocks in block allocators v1 and v2

Before the fix we used seq_len=0 for padding samples. This was later
translated to an empty attention_mask (since we don't have any tokens
that we should include in calculations) and in turn caused NaNs in
prompt attention (0 divided by 0). Those NaNs later got propagated to
kv-cache causing issues in flat_pa.

#256 breaks LoRA specific flow
which was handled through `is_profile_run` flag to distinguish warmup
and profile-run phase.

Introduces a new flag `is_lora_profile_run` to handle this LoRA specific
flow in profile-run.