Support FP8 INC in vLLM (#144)

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>

README_GAUDI.md

@@ -26,7 +26,8 @@ To verify that the Intel Gaudi software was correctly installed, run:
 ``` {.console}
 $ hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible
 $ apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core and habanalabs-thunk are installed
-$ pip list | habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml, habana-media-loader and habana_quantization_toolkit are installed
+$ pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed
+$ pip list | grep neural # verify that neural-compressor is installed
 ```
 
 Refer to [Intel Gaudi Software Stack

docs/source/getting_started/gaudi-installation.rst

@@ -26,7 +26,8 @@ To verify that the Intel Gaudi software was correctly installed, run:
 
    $ hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible
    $ apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core and habanalabs-thunk are installed
-   $ pip list | habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml, habana-media-loader and habana_quantization_toolkit are installed
+   $ pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed
+   $ pip list | grep neural # verify that neural_compressor is installed
 
 Refer to `Intel Gaudi Software Stack
 Verification <https://docs.habana.ai/en/latest/Installation_Guide/SW_Verification.html#platform-upgrade>`__

vllm/attention/backends/habana_attn.py

@@ -12,6 +12,8 @@
                                               AttentionMetadata, AttentionType)
 from vllm.attention.ops.habana_paged_attn import (HabanaPagedAttention,
                                                   HabanaPagedAttentionMetadata)
+from vllm.hpu import cache_ops
+from vllm.hpu.utils import Matmul, Softmax, VLLMKVCache
 from vllm.logger import init_logger
 
 logger = init_logger(__name__)
@@ -108,7 +110,7 @@ def __post_init__(self):
         self.attn_bias: Optional[torch.Tensor] = None
 
 
-class HabanaAttentionImpl(AttentionImpl):
+class HabanaAttentionImpl(AttentionImpl, torch.nn.Module):
     """
     If the input tensors contain prompt tokens, the layout is as follows:
     |<--------------- num_prefill_tokens ----------------->|
@@ -137,10 +139,16 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         max_seq_len: int = 4096,
     ) -> None:
+        super(AttentionImpl, self).__init__()
         self.kv_cache_dtype = kv_cache_dtype
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
+        self.matmul_qk = Matmul()
+        self.softmax = Softmax()
+        self.matmul_av = Matmul()
+        self.k_cache = VLLMKVCache()
+        self.v_cache = VLLMKVCache()
         self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
         self.sliding_window = sliding_window
         self.position_bias = None
@@ -204,9 +212,13 @@ def forward(
             # Reshape the input keys and values and store them in the cache.
             # If kv_cache is not provided, the new key and value tensors are
             # not cached. This happens during the initial memory profiling run.
-            HabanaPagedAttention.write_to_paged_cache(
-                key, value, key_cache, value_cache, attn_metadata.slot_mapping,
-                self.kv_cache_dtype, attn_metadata.is_prompt)
+            num_kv_cache_passes, num_slots_available, indices, offsets = \
+                cache_ops.prepare_to_cache(key_cache,
+                                           attn_metadata.slot_mapping)
+            key_cache = self.k_cache(key, key_cache, num_kv_cache_passes,
+                                     num_slots_available, indices, offsets)
+            value_cache = self.v_cache(value, value_cache, num_kv_cache_passes,
+                                       num_slots_available, indices, offsets)
 
         if attn_metadata.is_prompt:
             # Prompt run.
@@ -232,6 +244,9 @@ def forward(
                     attn_bias=attn_bias,
                     p=0.0,
                     scale=self.scale,
+                    matmul_qk_op=self.matmul_qk,
+                    softmax_op=self.softmax,
+                    matmul_av_op=self.matmul_av,
                 )
                 output = out.reshape(batch_size, seq_len, hidden_size)
             else:
@@ -255,7 +270,8 @@ def forward(
                 query, key_cache, value_cache, attn_metadata.block_tables,
                 attn_metadata.seq_lens_tensor, self.kv_cache_dtype,
                 self.num_kv_heads, self.scale, self.position_bias, k_scale,
-                v_scale)
+                v_scale, self.matmul_qk, self.softmax, self.matmul_av,
+                self.k_cache, self.v_cache)
         # Reshape the output tensor.
         return output.view(batch_size, seq_len, hidden_size)
 

vllm/attention/ops/habana_paged_attn.py

@@ -75,6 +75,11 @@ def forward_decode(
         alibi_slopes: Optional[torch.Tensor],
         k_scale: float,
         v_scale: float,
+        matmul_qk_op,
+        softmax_op,
+        matmul_av_op,
+        k_cache_cls,
+        v_cache_cls,
     ) -> torch.Tensor:
         block_size = value_cache.shape[1]
         return ops.paged_attention_v1(
@@ -88,6 +93,11 @@ def forward_decode(
             block_size,
             alibi_slopes,
             kv_cache_dtype,
+            matmul_qk_op,
+            softmax_op,
+            matmul_av_op,
+            k_cache_cls,
+            v_cache_cls,
         )
 
     @staticmethod

vllm/config.py

@@ -474,12 +474,13 @@ def _verify_args(self) -> None:
     def _verify_cache_dtype(self) -> None:
         if self.cache_dtype == "auto":
             pass
-        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2"):
+        elif self.cache_dtype in ("fp8", "fp8_e4m3", "fp8_e5m2", "fp8_inc"):
             logger.info(
                 "Using fp8 data type to store kv cache. It reduces the GPU "
                 "memory footprint and boosts the performance. "
                 "Meanwhile, it may cause accuracy drop without a proper "
-                "scaling factor")
+                "scaling factor. "
+                "Intel Gaudi (HPU) supports fp8 (using fp8_inc).")
         else:
             raise ValueError(f"Unknown kv cache dtype: {self.cache_dtype}")
 
@@ -600,11 +601,12 @@ class LoadConfig:
         ignore_patterns: The list of patterns to ignore when loading the model.
             Default to "original/**/*" to avoid repeated loading of llama's 
             checkpoints.
-            
+        device: Device on which weights are loaded.
     """
 
     load_format: Union[str, LoadFormat, "BaseModelLoader"] = LoadFormat.AUTO
     download_dir: Optional[str] = None
+    device: Optional[str] = None
     model_loader_extra_config: Optional[Union[str, dict]] = field(
         default_factory=dict)
     ignore_patterns: Optional[Union[List[str], str]] = None

vllm/engine/arg_utils.py

@@ -38,6 +38,7 @@ class EngineArgs:
     trust_remote_code: bool = False
     download_dir: Optional[str] = None
     load_format: str = 'auto'
+    weights_load_device: Optional[str] = None
     dtype: str = 'auto'
     kv_cache_dtype: str = 'auto'
     quantization_param_path: Optional[str] = None
@@ -205,6 +206,11 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
             'section for more information.\n'
             '* "bitsandbytes" will load the weights using bitsandbytes '
             'quantization.\n')
+        parser.add_argument("--weights-load-device",
+                            type=str,
+                            default=EngineArgs.weights_load_device,
+                            choices=["cuda", "neuron", "hpu", "cpu"],
+                            help='Device on which weights are loaded.')
         parser.add_argument(
             '--dtype',
             type=str,
@@ -223,11 +229,12 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
         parser.add_argument(
             '--kv-cache-dtype',
             type=str,
-            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
+            choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3', 'fp8_inc'],
             default=EngineArgs.kv_cache_dtype,
             help='Data type for kv cache storage. If "auto", will use model '
             'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
-            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
+            'ROCm (AMD GPU) supports fp8 (=fp8_e4m3). '
+            'Intel Gaudi (HPU) supports fp8 (using fp8_inc).')
         parser.add_argument(
             '--quantization-param-path',
             type=nullable_str,
@@ -835,9 +842,12 @@ def create_engine_config(self, ) -> EngineConfig:
             self.model_loader_extra_config[
                 "qlora_adapter_name_or_path"] = self.qlora_adapter_name_or_path
 
+        device = device_config.device if self.weights_load_device is None else \
+                 self.weights_load_device
         load_config = LoadConfig(
             load_format=self.load_format,
             download_dir=self.download_dir,
+            device=device,
             model_loader_extra_config=self.model_loader_extra_config,
             ignore_patterns=self.ignore_patterns,
         )

vllm/engine/llm_engine.py

@@ -182,7 +182,7 @@ def __init__(
             "download_dir=%r, load_format=%s, tensor_parallel_size=%d, "
             "pipeline_parallel_size=%d, "
             "disable_custom_all_reduce=%s, quantization=%s, "
-            "enforce_eager=%s, kv_cache_dtype=%s, "
+            "weights_load_device=%s, enforce_eager=%s, kv_cache_dtype=%s, "
             "quantization_param_path=%s, device_config=%s, "
             "decoding_config=%r, observability_config=%r, "
             "seed=%d, served_model_name=%s, use_v2_block_manager=%s, "
@@ -206,6 +206,7 @@ def __init__(
             parallel_config.pipeline_parallel_size,
             parallel_config.disable_custom_all_reduce,
             model_config.quantization,
+            load_config.device,
             model_config.enforce_eager,
             cache_config.cache_dtype,
             model_config.quantization_param_path,
@@ -853,6 +854,9 @@ def _process_model_outputs(
             request_outputs.append(request_output)
         return request_outputs
 
+    def finish_measurements(self):
+        self.model_executor.finish_measurements()
+
     def step(self) -> List[Union[RequestOutput, EmbeddingRequestOutput]]:
         """Performs one decoding iteration and returns newly generated results.
 

vllm/entrypoints/llm.py

@@ -173,6 +173,9 @@ def set_tokenizer(
             self.llm_engine.tokenizer.tokenizer = get_cached_tokenizer(
                 tokenizer)
 
+    def finish_measurements(self):
+        self.llm_engine.finish_measurements()
+
     @overload  # LEGACY: single (prompt + optional token ids)
     def generate(
         self,

vllm/executor/habana_executor.py

@@ -90,6 +90,9 @@ def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks) -> None:
         msg = f"init_cache_engine took {cache_init_m.get_summary_string()}"
         logger.info(msg)
 
+    def finish_measurements(self):
+        self.driver_worker.finish_measurements()
+
     def execute_model(
             self,
             execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:
@@ -180,6 +183,12 @@ def check_health(self) -> None:
         # it's running.
         return
 
+    def shutdown(self) -> None:
+        self.driver_worker.shutdown_inc()
+
+    def __del__(self):
+        self.shutdown()
+
 
 class HabanaExecutorAsync(HabanaExecutor, ExecutorAsyncBase):
 

vllm/executor/ray_habana_executor.py

@@ -237,6 +237,9 @@ def _driver_execute_model(
         return self.driver_worker.execute_method("execute_model",
                                                  execute_model_req)
 
+    def finish_measurements(self):
+        self._run_workers("finish_measurements")
+
     def execute_model(
             self,
             execute_model_req: ExecuteModelRequest) -> List[SamplerOutput]:

vllm/hpu/cache_ops.py

@@ -43,6 +43,37 @@ def reshape_and_cache(key,
             value[start_idx:end_idx])
 
 
+def prepare_to_cache(cache, slot_mapping):
+    num_blocks = cache.size(0)
+    block_size = cache.size(1)
+    slot_mapping = slot_mapping.flatten()
+    indices = torch.div(slot_mapping, block_size, rounding_mode="floor")
+    offsets = torch.fmod(slot_mapping, block_size)
+    num_slots_requested = slot_mapping.size(0)
+    num_slots_available = num_blocks * block_size
+    # NOTE(kzawora): HPU PT bridge crashes with
+    # RuntimeError: Invalid inputs for scatter_nd_onnx
+    # on index_put when num_slots_requested > num_slots_available.
+    # This case might occur when we have little kv cache blocks and
+    # lots of padding, or are doing warmup.
+    # This loop is a workaround for this issue. Please remove it
+    # once key_cache.index_put_(indices, offsets), key) works.
+    num_kv_cache_passes = torch.div(num_slots_requested,
+                                    num_slots_available).ceil().int().item()
+
+    return num_kv_cache_passes, num_slots_available, indices, offsets
+
+
+def insert_or_update_cache(input, cache, num_kv_cache_passes,
+                           num_slots_available, block_indices, block_offsets):
+    for i in range(num_kv_cache_passes):
+        start_idx = i * num_slots_available
+        end_idx = (i + 1) * num_slots_available
+        cache.index_put_((block_indices[start_idx:end_idx],
+                          block_offsets[start_idx:end_idx]),
+                         input[start_idx:end_idx])
+
+
 def swap_blocks(src, dst, block_mapping):
     index_src = torch.zeros((1, ), dtype=torch.int32, device=src.device)
     index_dst = torch.zeros((1, ), dtype=torch.int32, device=dst.device)