optimum-rbln 0.7.3a4__py3-none-any.whl → 0.7.3a6__py3-none-any.whl

optimum/rbln/__init__.py

@@ -78,9 +78,13 @@ _import_structure = {
         "RBLNAutoencoderKL",
         "RBLNControlNetModel",
         "RBLNPriorTransformer",
+        "RBLNKandinskyV22CombinedPipeline",
+        "RBLNKandinskyV22Img2ImgCombinedPipeline",
         "RBLNKandinskyV22InpaintCombinedPipeline",
         "RBLNKandinskyV22InpaintPipeline",
+        "RBLNKandinskyV22Img2ImgPipeline",
         "RBLNKandinskyV22PriorPipeline",
+        "RBLNKandinskyV22Pipeline",
         "RBLNStableDiffusionPipeline",
         "RBLNStableDiffusionXLPipeline",
         "RBLNUNet2DConditionModel",
@@ -107,8 +111,12 @@ if TYPE_CHECKING:
         RBLNAutoencoderKL,
         RBLNControlNetModel,
         RBLNDiffusionMixin,
+        RBLNKandinskyV22CombinedPipeline,
+        RBLNKandinskyV22Img2ImgCombinedPipeline,
+        RBLNKandinskyV22Img2ImgPipeline,
         RBLNKandinskyV22InpaintCombinedPipeline,
         RBLNKandinskyV22InpaintPipeline,
+        RBLNKandinskyV22Pipeline,
         RBLNKandinskyV22PriorPipeline,
         RBLNMultiControlNetModel,
         RBLNPriorTransformer,

optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.3a4'
-__version_tuple__ = version_tuple = (0, 7, 3, 'a4')
+__version__ = version = '0.7.3a6'
+__version_tuple__ = version_tuple = (0, 7, 3, 'a6')

optimum/rbln/diffusers/__init__.py

@@ -24,9 +24,13 @@ ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES["optimum.rbln"])
 
 _import_structure = {
     "pipelines": [
+        "RBLNKandinskyV22CombinedPipeline",
+        "RBLNKandinskyV22Img2ImgCombinedPipeline",
         "RBLNKandinskyV22InpaintCombinedPipeline",
         "RBLNKandinskyV22InpaintPipeline",
+        "RBLNKandinskyV22Img2ImgPipeline",
         "RBLNKandinskyV22PriorPipeline",
+        "RBLNKandinskyV22Pipeline",
         "RBLNStableDiffusionPipeline",
         "RBLNStableDiffusionXLPipeline",
         "RBLNStableDiffusionImg2ImgPipeline",
@@ -66,8 +70,12 @@ if TYPE_CHECKING:
         RBLNVQModel,
     )
     from .pipelines import (
+        RBLNKandinskyV22CombinedPipeline,
+        RBLNKandinskyV22Img2ImgCombinedPipeline,
+        RBLNKandinskyV22Img2ImgPipeline,
         RBLNKandinskyV22InpaintCombinedPipeline,
         RBLNKandinskyV22InpaintPipeline,
+        RBLNKandinskyV22Pipeline,
         RBLNKandinskyV22PriorPipeline,
         RBLNMultiControlNetModel,
         RBLNStableDiffusion3Img2ImgPipeline,

optimum/rbln/diffusers/models/autoencoders/vq_model.py

@@ -90,9 +90,17 @@ class RBLNVQModel(RBLNModel):
         model_config: "PretrainedConfig",
         rbln_kwargs: Dict[str, Any] = {},
     ) -> RBLNConfig:
-        batch_size = rbln_kwargs.get("batch_size") or 1
-        height = rbln_kwargs.get("img_height") or 512
-        width = rbln_kwargs.get("img_width") or 512
+        batch_size = rbln_kwargs.get("batch_size")
+        if batch_size is None:
+            batch_size = 1
+
+        height = rbln_kwargs.get("img_height")
+        if height is None:
+            height = 512
+
+        width = rbln_kwargs.get("img_width")
+        if width is None:
+            width = 512
 
         if hasattr(model_config, "block_out_channels"):
             scale_factor = 2 ** (len(model_config.block_out_channels) - 1)

optimum/rbln/diffusers/models/unets/unet_2d_condition.py

@@ -176,15 +176,22 @@ class RBLNUNet2DConditionModel(RBLNModel):
             raise ValueError("Both image height and image width must be given or not given")
         elif image_size[0] is None and image_size[1] is None:
             if rbln_config["img2img_pipeline"]:
-                # In case of img2img, sample size of unet is determined by vae encoder.
-                vae_sample_size = pipe.vae.config.sample_size
-                if isinstance(vae_sample_size, int):
-                    sample_size = vae_sample_size // scale_factor
-                else:
-                    sample_size = (
-                        vae_sample_size[0] // scale_factor,
-                        vae_sample_size[1] // scale_factor,
+                if hasattr(pipe, "vae"):
+                    # In case of img2img, sample size of unet is determined by vae encoder.
+                    vae_sample_size = pipe.vae.config.sample_size
+                    if isinstance(vae_sample_size, int):
+                        sample_size = vae_sample_size // scale_factor
+                    else:
+                        sample_size = (
+                            vae_sample_size[0] // scale_factor,
+                            vae_sample_size[1] // scale_factor,
+                        )
+                elif hasattr(pipe, "movq"):
+                    logger.warning(
+                        "RBLN config 'img_height' and 'img_width' should have been provided for this pipeline. "
+                        "Both variable will be set 512 by default."
                     )
+                    sample_size = (512 // scale_factor, 512 // scale_factor)
             else:
                 sample_size = pipe.unet.config.sample_size
         else:

optimum/rbln/diffusers/pipelines/__init__.py

@@ -26,9 +26,13 @@ _import_structure = {
         "RBLNStableDiffusionXLControlNetPipeline",
     ],
     "kandinsky2_2": [
+        "RBLNKandinskyV22CombinedPipeline",
+        "RBLNKandinskyV22Img2ImgCombinedPipeline",
         "RBLNKandinskyV22InpaintCombinedPipeline",
         "RBLNKandinskyV22InpaintPipeline",
+        "RBLNKandinskyV22Img2ImgPipeline",
         "RBLNKandinskyV22PriorPipeline",
+        "RBLNKandinskyV22Pipeline",
     ],
     "stable_diffusion": [
         "RBLNStableDiffusionImg2ImgPipeline",
@@ -55,8 +59,12 @@ if TYPE_CHECKING:
         RBLNStableDiffusionXLControlNetPipeline,
     )
     from .kandinsky2_2 import (
+        RBLNKandinskyV22CombinedPipeline,
+        RBLNKandinskyV22Img2ImgCombinedPipeline,
+        RBLNKandinskyV22Img2ImgPipeline,
         RBLNKandinskyV22InpaintCombinedPipeline,
         RBLNKandinskyV22InpaintPipeline,
+        RBLNKandinskyV22Pipeline,
         RBLNKandinskyV22PriorPipeline,
     )
     from .stable_diffusion import (

optimum/rbln/diffusers/pipelines/kandinsky2_2/__init__.py

@@ -12,6 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .pipeline_kandinsky2_2_combined import RBLNKandinskyV22InpaintCombinedPipeline
+from .pipeline_kandinsky2_2 import RBLNKandinskyV22Pipeline
+from .pipeline_kandinsky2_2_combined import (
+    RBLNKandinskyV22CombinedPipeline,
+    RBLNKandinskyV22Img2ImgCombinedPipeline,
+    RBLNKandinskyV22InpaintCombinedPipeline,
+)
+from .pipeline_kandinsky2_2_img2img import RBLNKandinskyV22Img2ImgPipeline
 from .pipeline_kandinsky2_2_inpaint import RBLNKandinskyV22InpaintPipeline
 from .pipeline_kandinsky2_2_prior import RBLNKandinskyV22PriorPipeline

optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py

@@ -0,0 +1,25 @@
+# Copyright 2024 Rebellions Inc.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from diffusers import KandinskyV22Pipeline
+
+from ...modeling_diffusers import RBLNDiffusionMixin
+
+
+class RBLNKandinskyV22Pipeline(RBLNDiffusionMixin, KandinskyV22Pipeline):
+    original_class = KandinskyV22Pipeline
+    _submodules = ["unet", "movq"]
+
+    def get_compiled_image_size(self):
+        return self.movq.image_size

optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py

@@ -14,6 +14,8 @@
 
 from diffusers import (
     DDPMScheduler,
+    KandinskyV22CombinedPipeline,
+    KandinskyV22Img2ImgCombinedPipeline,
     KandinskyV22InpaintCombinedPipeline,
     PriorTransformer,
     UnCLIPScheduler,
@@ -28,10 +30,114 @@ from transformers import (
 )
 
 from ...modeling_diffusers import RBLNDiffusionMixin
+from .pipeline_kandinsky2_2 import RBLNKandinskyV22Pipeline
+from .pipeline_kandinsky2_2_img2img import RBLNKandinskyV22Img2ImgPipeline
 from .pipeline_kandinsky2_2_inpaint import RBLNKandinskyV22InpaintPipeline
 from .pipeline_kandinsky2_2_prior import RBLNKandinskyV22PriorPipeline
 
 
+class RBLNKandinskyV22CombinedPipeline(RBLNDiffusionMixin, KandinskyV22CombinedPipeline):
+    original_class = KandinskyV22CombinedPipeline
+    _connected_classes = {"prior_pipe": RBLNKandinskyV22PriorPipeline, "decoder_pipe": RBLNKandinskyV22Pipeline}
+    _submodules = ["prior_pipe", "decoder_pipe"]
+    _prefix = {"prior_pipe": "prior_"}
+
+    def __init__(
+        self,
+        unet: "UNet2DConditionModel",
+        scheduler: "DDPMScheduler",
+        movq: "VQModel",
+        prior_prior: "PriorTransformer",
+        prior_image_encoder: "CLIPVisionModelWithProjection",
+        prior_text_encoder: "CLIPTextModelWithProjection",
+        prior_tokenizer: "CLIPTokenizer",
+        prior_scheduler: "UnCLIPScheduler",
+        prior_image_processor: "CLIPImageProcessor",
+    ):
+        RBLNDiffusionMixin.__init__(self)
+        super(KandinskyV22CombinedPipeline, self).__init__()
+
+        self.register_modules(
+            unet=unet,
+            scheduler=scheduler,
+            movq=movq,
+            prior_prior=prior_prior,
+            prior_image_encoder=prior_image_encoder,
+            prior_text_encoder=prior_text_encoder,
+            prior_tokenizer=prior_tokenizer,
+            prior_scheduler=prior_scheduler,
+            prior_image_processor=prior_image_processor,
+        )
+
+        self.prior_pipe = RBLNKandinskyV22PriorPipeline(
+            prior=prior_prior,
+            image_encoder=prior_image_encoder,
+            text_encoder=prior_text_encoder,
+            tokenizer=prior_tokenizer,
+            scheduler=prior_scheduler,
+            image_processor=prior_image_processor,
+        )
+        self.decoder_pipe = RBLNKandinskyV22Pipeline(
+            unet=unet,
+            scheduler=scheduler,
+            movq=movq,
+        )
+
+    def get_compiled_image_size(self):
+        return self.movq.image_size
+
+
+class RBLNKandinskyV22Img2ImgCombinedPipeline(RBLNDiffusionMixin, KandinskyV22Img2ImgCombinedPipeline):
+    original_class = KandinskyV22Img2ImgCombinedPipeline
+    _connected_classes = {"prior_pipe": RBLNKandinskyV22PriorPipeline, "decoder_pipe": RBLNKandinskyV22Img2ImgPipeline}
+    _submodules = ["prior_pipe", "decoder_pipe"]
+    _prefix = {"prior_pipe": "prior_"}
+
+    def __init__(
+        self,
+        unet: "UNet2DConditionModel",
+        scheduler: "DDPMScheduler",
+        movq: "VQModel",
+        prior_prior: "PriorTransformer",
+        prior_image_encoder: "CLIPVisionModelWithProjection",
+        prior_text_encoder: "CLIPTextModelWithProjection",
+        prior_tokenizer: "CLIPTokenizer",
+        prior_scheduler: "UnCLIPScheduler",
+        prior_image_processor: "CLIPImageProcessor",
+    ):
+        RBLNDiffusionMixin.__init__(self)
+        super(KandinskyV22Img2ImgCombinedPipeline, self).__init__()
+
+        self.register_modules(
+            unet=unet,
+            scheduler=scheduler,
+            movq=movq,
+            prior_prior=prior_prior,
+            prior_image_encoder=prior_image_encoder,
+            prior_text_encoder=prior_text_encoder,
+            prior_tokenizer=prior_tokenizer,
+            prior_scheduler=prior_scheduler,
+            prior_image_processor=prior_image_processor,
+        )
+
+        self.prior_pipe = RBLNKandinskyV22PriorPipeline(
+            prior=prior_prior,
+            image_encoder=prior_image_encoder,
+            text_encoder=prior_text_encoder,
+            tokenizer=prior_tokenizer,
+            scheduler=prior_scheduler,
+            image_processor=prior_image_processor,
+        )
+        self.decoder_pipe = RBLNKandinskyV22Img2ImgPipeline(
+            unet=unet,
+            scheduler=scheduler,
+            movq=movq,
+        )
+
+    def get_compiled_image_size(self):
+        return self.movq.image_size
+
+
 class RBLNKandinskyV22InpaintCombinedPipeline(RBLNDiffusionMixin, KandinskyV22InpaintCombinedPipeline):
     original_class = KandinskyV22InpaintCombinedPipeline
     _connected_classes = {"prior_pipe": RBLNKandinskyV22PriorPipeline, "decoder_pipe": RBLNKandinskyV22InpaintPipeline}

optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py

@@ -0,0 +1,25 @@
+# Copyright 2024 Rebellions Inc.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from diffusers import KandinskyV22Img2ImgPipeline
+
+from ...modeling_diffusers import RBLNDiffusionMixin
+
+
+class RBLNKandinskyV22Img2ImgPipeline(RBLNDiffusionMixin, KandinskyV22Img2ImgPipeline):
+    original_class = KandinskyV22Img2ImgPipeline
+    _submodules = ["unet", "movq"]
+
+    def get_compiled_image_size(self):
+        return self.movq.image_size

optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpaint.py

@@ -20,3 +20,6 @@ from ...modeling_diffusers import RBLNDiffusionMixin
 class RBLNKandinskyV22InpaintPipeline(RBLNDiffusionMixin, KandinskyV22InpaintPipeline):
     original_class = KandinskyV22InpaintPipeline
     _submodules = ["unet", "movq"]
+
+    def get_compiled_image_size(self):
+        return self.movq.image_size

optimum/rbln/modeling_base.py

@@ -282,15 +282,6 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
             **kwargs,
         )
 
-    @classmethod
-    def _check_compiled_models(
-        cls, compiled_models: Dict[str, rebel.RBLNCompiledModel], rbln_config: RBLNConfig, config: "PretrainedConfig"
-    ):
-        # check compiled model can create runtimes.
-        # this logic currently only works in LLM
-        # fail when LLM model using Paged Attention can't guarantee max sequence length
-        pass
-
     @classmethod
     def _from_compiled_models(
         cls,
@@ -305,8 +296,6 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
         if isinstance(model_save_dir, str):
             model_save_dir = Path(model_save_dir)
 
-        cls._check_compiled_models(compiled_models=rbln_compiled_models, rbln_config=rbln_config, config=config)
-
         # FIXME:: Should we convert it?
         compiled_model_names = [cfg.compiled_model_name for cfg in rbln_config.compile_cfgs]
         rbln_compiled_models = [rbln_compiled_models[cm_name] for cm_name in compiled_model_names]

optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py

@@ -98,9 +98,9 @@ class RBLNRuntimeModel(RBLNPytorchRuntime):
         """
 
         NO_BLOCKS_ERROR = (
-            "No memory blocks are available for allocation."
-            "The generate() API cannot complete this inference task because Paged Attention is not fully supported by optimum-rbln."
-            "This is supported by vllm-rbln (see: https://docs.rbln.ai/software/model_serving/vllm_support/vllm-rbln.html)."
+            "No memory blocks are available for allocation. "
+            "The generate() API cannot complete this inference task because Paged Attention is not fully supported by optimum-rbln. "
+            "This is supported by vllm-rbln (see: https://docs.rbln.ai/software/model_serving/vllm_support/vllm-rbln.html). "
             "Using vllm-rbln should fix this issue and enhance inference performance."
         )
 
@@ -575,59 +575,58 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
         nbits_per_param: int,
         n_model_params: int,
     ) -> int:
-        num_attention_heads = getattr(config, "n_head", None) or getattr(config, "num_attention_heads")
-        num_layers = getattr(config, "n_layer", None) or getattr(config, "num_hidden_layers")
-        head_dim = getattr(config, "head_dim", None) or config.hidden_size // num_attention_heads
-        vocab_size = config.vocab_size
-        hidden_size = getattr(config, "n_embd", None) or getattr(config, "hidden_size")
-        num_key_value_heads = getattr(config, "num_key_value_heads", None) or num_attention_heads
-
-        TARGET_DRAM_LIMIT = int(tensor_parallel_size * 15.7 * 2**30)  # 16GB # TODO(jongho): 더 정확한 값
-
         def align(x: int, nbytes: int) -> int:
             return int(math.ceil(x / nbytes) * nbytes)
 
         def align_2MB(x: int) -> int:
             return align(x, 2 * 1024 * 1024)
 
-        def get_kernel_size() -> int:
-            # TODO: Implement
-            lm_heads_params = align(vocab_size, 64) * hidden_size
-            lm_heads_nbytes = (
-                align_2MB(lm_heads_params * nbits_per_param // 8 / tensor_parallel_size) * tensor_parallel_size
-            )
+        num_attention_heads = getattr(config, "n_head", None) or getattr(config, "num_attention_heads")
+        num_layers = getattr(config, "n_layer", None) or getattr(config, "num_hidden_layers")
+        head_dim = getattr(config, "head_dim", None) or config.hidden_size // num_attention_heads
+        vocab_size = config.vocab_size
+        hidden_size = getattr(config, "n_embd", None) or getattr(config, "hidden_size")
+        num_key_value_heads = getattr(config, "num_key_value_heads", None) or num_attention_heads
 
-            params = n_model_params - lm_heads_params
-            layer_nbytes = (
-                align_2MB(params * nbits_per_param // 8 / num_layers / tensor_parallel_size)
-                * num_layers
-                * tensor_parallel_size
-            )
+        # TODO(jongho): Update if target npu is REBEL.
+        ATOM_DRAM_NBYTES = 16 * 2**30
+        ATOM_SYS_DRAM_NBYTES = 288 * 2**20
+        available_dram = tensor_parallel_size * (ATOM_DRAM_NBYTES - ATOM_SYS_DRAM_NBYTES)
 
-            return layer_nbytes + lm_heads_nbytes
+        # Get estimated kernel size (approximated)
+        lm_heads_params = align(vocab_size, 64) * hidden_size
+        lm_heads_nbytes = (
+            align_2MB(lm_heads_params * nbits_per_param // 8 / tensor_parallel_size) * tensor_parallel_size
+        )
+        params = n_model_params - lm_heads_params
+        layer_nbytes = (
+            align_2MB(params * nbits_per_param // 8 / num_layers / tensor_parallel_size)
+            * num_layers
+            * tensor_parallel_size
+        )
+        kernel_size = layer_nbytes + lm_heads_nbytes
 
-        available_dram = TARGET_DRAM_LIMIT - get_kernel_size()
+        available_dram -= kernel_size
 
-        buffer = 2**30  # 1GB
-        if tensor_parallel_size <= 2:
+        # TODO: Accurate buffer estimation
+        buffer = 2**30  # 1GB Buffer
+        if tensor_parallel_size <= 4:
             buffer /= 4
 
         available_dram -= buffer
 
-        def get_nbytes_per_block() -> int:
-            return (
-                align_2MB(
-                    kvcache_block_size
-                    * head_dim
-                    * math.ceil(num_key_value_heads / tensor_parallel_size)  # Shard
-                    * 2  # (fp16)
-                )
-                * num_layers
-                * 2  # (k, v)
-                * tensor_parallel_size
+        # Estimate nbytes per a single kvcache block
+        nbytes_per_block = (
+            align_2MB(
+                kvcache_block_size
+                * head_dim
+                * math.ceil(num_key_value_heads / tensor_parallel_size)  # Shard
+                * 2  # (fp16)
             )
-
-        nbytes_per_block = get_nbytes_per_block()
+            * num_layers
+            * 2  # (k, v)
+            * tensor_parallel_size
+        )
         n_blocks = available_dram // nbytes_per_block
 
         return n_blocks, nbytes_per_block
@@ -685,27 +684,28 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
             else:
                 rbln_kvcache_block_size = rbln_kvcache_partition_len
 
-        max_num_blocks, nbytes_per_block = cls.get_maximum_num_blocks(
-            config=model_config,
-            tensor_parallel_size=rbln_kwargs.get("tensor_parallel_size", 1),
-            kvcache_block_size=rbln_kvcache_block_size,
-            nbits_per_param=16 if rbln_quantization is None else 4,  # TODO(jongho): FIX Ad-hoc
-            n_model_params=rbln_kwargs["n_model_params"],
-        )
-        model_num_blocks = (rbln_max_seq_len // rbln_kvcache_block_size) * rbln_batch_size
-        rbln_kvcache_num_blocks = min(model_num_blocks, max_num_blocks)
+        rbln_kvcache_num_blocks = (rbln_max_seq_len // rbln_kvcache_block_size) * rbln_batch_size
+        if rbln_attn_impl == "flash_attn":
+            max_num_blocks, _ = cls.get_maximum_num_blocks(
+                config=model_config,
+                tensor_parallel_size=rbln_kwargs.get("tensor_parallel_size", 1),
+                kvcache_block_size=rbln_kvcache_block_size,
+                nbits_per_param=16 if rbln_quantization is None else 4,  # TODO(jongho): FIX Ad-hoc
+                n_model_params=rbln_kwargs["n_model_params"],
+            )
+            rbln_kvcache_num_blocks = min(rbln_kvcache_num_blocks, max_num_blocks)
 
-        required_blocks = rbln_max_seq_len // rbln_kvcache_block_size + 1
-        if rbln_kvcache_num_blocks < required_blocks:
-            rbln_kvcache_num_blocks = required_blocks
+            required_blocks = rbln_max_seq_len // rbln_kvcache_block_size + 1
+            if rbln_kvcache_num_blocks < required_blocks:
+                rbln_kvcache_num_blocks = required_blocks
 
-        logger.info(f"[KVCache] Compiling with num_blocks: {rbln_kvcache_num_blocks}")
+            logger.info(f"[KVCache] Compiling with num_blocks: {rbln_kvcache_num_blocks}")
 
-        if rbln_kvcache_num_blocks < rbln_batch_size:
-            raise RuntimeError(
-                f"Batch size ({rbln_batch_size}) exceeds available KV cache blocks ({rbln_kvcache_num_blocks}). "
-                "Ensure the number of blocks is at least equal to the batch size."
-            )
+            if rbln_kvcache_num_blocks < rbln_batch_size:
+                raise RuntimeError(
+                    f"Batch size ({rbln_batch_size}) exceeds available KV cache blocks ({rbln_kvcache_num_blocks}). "
+                    "Ensure the number of blocks is at least equal to the batch size."
+                )
 
         num_attention_heads = getattr(model_config, "n_head", None) or getattr(model_config, "num_attention_heads")
         num_key_value_heads = getattr(model_config, "num_key_value_heads", None) or num_attention_heads
@@ -805,9 +805,6 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
                 "kvcache_block_size": rbln_kvcache_block_size,
                 "attn_impl": rbln_attn_impl,
                 "kvcache_num_blocks": rbln_kvcache_num_blocks,
-                "model_num_blocks": model_num_blocks,
-                "max_num_blocks": max_num_blocks,
-                "nbytes_per_block": nbytes_per_block,
             }
         )
 

optimum/rbln/transformers/models/phi/phi_architecture.py

@@ -92,7 +92,7 @@ class PhiLayer(DecoderOnlyLayer):
 
         hidden_states = self.get_pre_attention_layernorm()(hidden_states)
 
-        attn_outputs, present_key_values = self.self_attn(
+        attn_output = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
             seq_positions=seq_positions,
@@ -104,9 +104,9 @@ class PhiLayer(DecoderOnlyLayer):
 
         feed_forward_hidden_states = self._original_mod.mlp(hidden_states)
 
-        hidden_states = attn_outputs + feed_forward_hidden_states + residual
+        hidden_states = attn_output + feed_forward_hidden_states + residual
 
-        return hidden_states, present_key_values
+        return hidden_states
 
 
 class PhiModel(DecoderOnlyModel):

{optimum_rbln-0.7.3a4.dist-info → optimum_rbln-0.7.3a6.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.3a4
+Version: 0.7.3a6
 Summary: Optimum RBLN is the interface between the Hugging Face Transformers and Diffusers libraries and RBLN accelerators. It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
 Project-URL: Homepage, https://rebellions.ai
 Project-URL: Documentation, https://docs.rbln.ai

{optimum_rbln-0.7.3a4.dist-info → optimum_rbln-0.7.3a6.dist-info}/RECORD

@@ -1,31 +1,33 @@
-optimum/rbln/__init__.py,sha256=eHi15YM3989AcX52jka9rUmgAtlp1PHqMNwBEdOfuu8,6554
-optimum/rbln/__version__.py,sha256=MLlg_138GxyhciEP0ZB5dPN8vriXkicRnaZiwqygxOY,519
+optimum/rbln/__init__.py,sha256=ZDzXcl-oAcYJhKjJMpotjbTih9awo7HzUb6T3MUEP6Q,6894
+optimum/rbln/__version__.py,sha256=9voT1MrnPHKvqTeiZK8bNEZcPseZOq7N_U5etptnmTE,519
 optimum/rbln/modeling.py,sha256=nJsAs5zs--VVOYGFjYNpqfxYIemJIK4Lr0WEzlDLdP0,8390
-optimum/rbln/modeling_base.py,sha256=Ow73GVJF1N5cDFO8_rgirtGj1wC-cXBDyqXHW5PCybA,22270
+optimum/rbln/modeling_base.py,sha256=dNCL-BhrWCpuOVkZaj8-MW567Tf4lLo3p3Z3ldjWJfU,21779
 optimum/rbln/modeling_config.py,sha256=7104bxmrvKW4Q6XTruQayiIGl8GHDFmPkJ3cknMIInE,11335
-optimum/rbln/diffusers/__init__.py,sha256=pOyoXv3-JRzTBSwPKbgLS9H6F2K9dJdReEmpGhcLQYU,3283
+optimum/rbln/diffusers/__init__.py,sha256=Hq87CbtiCy85YmK2SB-OmUyfv77oe3j4bsTenTRnu6w,3623
 optimum/rbln/diffusers/modeling_diffusers.py,sha256=zqVNgH9oeOx2iNE7VsW_FinVf4s6G5Idyh4TKz7XJJg,21116
 optimum/rbln/diffusers/models/__init__.py,sha256=mkCvJyH1KcwrsUvYSq_bVC79oOfyqtBSFDyPS1_48wA,1478
 optimum/rbln/diffusers/models/controlnet.py,sha256=EM_HlzCdaZdnnK0oGpY2fQeigPqHhlwh4NHCzlmoumI,10512
 optimum/rbln/diffusers/models/autoencoders/__init__.py,sha256=dg17ZTUsiqTcbIaEE4fqew9uRbao0diQ21PXvRKIqKg,679
 optimum/rbln/diffusers/models/autoencoders/autoencoder_kl.py,sha256=rCbC32bJnfXtsLdVvNVVHpRAkCYy6jeCSwIZ-JSReWk,9220
 optimum/rbln/diffusers/models/autoencoders/vae.py,sha256=gB9HR7Bf7wpIXLv-Js4Pc3oyWRlqEe4cms4sI2AJicY,4380
-optimum/rbln/diffusers/models/autoencoders/vq_model.py,sha256=GunIau02_-lodYZBzd0ktJSNRT5axEFIZxSAfj2Mlyo,5974
+optimum/rbln/diffusers/models/autoencoders/vq_model.py,sha256=b36QqPbayjApKivceQVVyQxHyR1ZOZ1ffuGgdALEPTQ,6117
 optimum/rbln/diffusers/models/transformers/__init__.py,sha256=V8rSR7WzHs-i8Cwb_MNxhY2NFbwPgxu24vGtkwl-6tk,706
 optimum/rbln/diffusers/models/transformers/prior_transformer.py,sha256=VG9cQo-_eppDvQSW1q1euAGBt1socUHetN_fIN2u1iU,6169
 optimum/rbln/diffusers/models/transformers/transformer_sd3.py,sha256=n_krmMgiRxWrG--567PNpk58EG_X7x7H4gidIkRvwjo,7308
 optimum/rbln/diffusers/models/unets/__init__.py,sha256=MaICuK9CWjgzejXy8y2NDrphuEq1rkzanF8u45k6O5I,655
-optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=xHnBzFrm7aNaolxrsotbjo9GkbNiNdTleXQoeqGLlhg,15540
-optimum/rbln/diffusers/pipelines/__init__.py,sha256=DAsM4eNks3hEY-bsUKSxRKmgwUWDGDlw82gfplSOdO8,2800
+optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=QIjVWQQf8KBn5rU7lvipdm3gNBxZl7l6HCAj7p5FjLU,15977
+optimum/rbln/diffusers/pipelines/__init__.py,sha256=5KLZ5LrpMzBya2e_3_PvEoPwG24U8JMexfw_ygZREKc,3140
 optimum/rbln/diffusers/pipelines/controlnet/__init__.py,sha256=n1Ef22TSeax-kENi_d8K6wGGHSNEo9QkUeygELHgcao,983
 optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py,sha256=JWKtnZYBIfgmbAo0SLFIvHBQCv2BPSFNvpcdjG4GUOY,4113
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet.py,sha256=dGdw5cwJLS4CLv6IHskk5ZCcPgS7UDuHKbfOZ8ojNUs,35187
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py,sha256=7xCiXrH4ToCTHohVGFXqO7_f9G8HShYaHgZxoMZARkQ,33664
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py,sha256=Gzt2wg4dgFg0TV3Bu0cs8Xru3wVrxWUxxgciwZ-QKLE,44755
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py,sha256=RfwxNX_zQWFtvvFQJ5bt3qtHbdYdQV_3XLHm9WYCKOs,46084
-optimum/rbln/diffusers/pipelines/kandinsky2_2/__init__.py,sha256=YFqA76_XiMNxPwqotbHug2kd7jCbOXOu5NlxG2hbaVs,808
-optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py,sha256=9szfe1NvOr1mgDnSPZvBGq1b65RElUrqLVhuErY3Dmw,2962
-optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpaint.py,sha256=WxBbHAZSAKDSWhFerPvUlIhhWEsejW4NmhwmWX-_b54,856
+optimum/rbln/diffusers/pipelines/kandinsky2_2/__init__.py,sha256=I4YQq2HfA3xONbWsdJ870IEJPyLWeCDDG-UCJsu9YO8,1035
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py,sha256=aNFGOjth8tDvPrjYLbRWrkHr6p-8AFgcQx1Qay1fw70,904
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py,sha256=unqFDviA7dnx0yuo8L8tXVj2mjFYCPm7C9dcpdWBICc,6882
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py,sha256=fEs-WgJqWs5zvuCkKb7MuZokH9Mi6q-0DOEKxzfWxzo,932
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpaint.py,sha256=Ad2ZYCXaMiYpB0mz-8X1CGhILxrVbt7rRIXt6IPwYBM,932
 optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py,sha256=Mf7tzrXetwCgt7LuXfkX-CX1hltLgNZdwF9bHxAbDJM,874
 optimum/rbln/diffusers/pipelines/stable_diffusion/__init__.py,sha256=gz6CbP4T6w8XH3PIGRIJXTmKFsChJIkwcAEAsiR5Ydg,830
 optimum/rbln/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py,sha256=DgRLzO9HxtgE1jICmHoHaqeVXM4Ih-5uo2JqNMAPMcc,876
@@ -60,7 +62,7 @@ optimum/rbln/transformers/models/clip/__init__.py,sha256=H9vuBwrmFO0-CqZhXUrKF-u
 optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=NiSm7bHs4SReHDUr53BBWSX0Y8bkKOeUSpsBDrp8YDw,6628
 optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=pDogsdpJKKB5rqnVFrRjwfhUvOSV-jZ3oARMsqSvOOQ,665
 optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=7OIKteJLKNxOLOg0w3lLOM7TxZovQn4jkglI9wRkrtQ,40609
-optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=W9HnxJoTz78Wc4X5Q3sMSHhMTSa7-9uQCFlnqNVozvA,38932
+optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=uGdPGcFrWm2gAwFLjfBiALwFsl49VGCReVi4NUfOPxM,38898
 optimum/rbln/transformers/models/dpt/__init__.py,sha256=gP1tkR3XMNlHq1GT87ugIVvb2o_1eAUg1JaniXjy1Lw,651
 optimum/rbln/transformers/models/dpt/modeling_dpt.py,sha256=ZsS2SOiqcA4azULB-WFEMQZbgIoOyVUKqVKqrw_tWzA,3430
 optimum/rbln/transformers/models/exaone/__init__.py,sha256=zYH_5tVa8-juEdsOIky7I33WSC3Zuhoq1upI0OHYeVw,859
@@ -85,7 +87,7 @@ optimum/rbln/transformers/models/mistral/mistral_architecture.py,sha256=_aU8TE_t
 optimum/rbln/transformers/models/mistral/modeling_mistral.py,sha256=7nrddoBIHf8S12LZWBUpotnvG3gND11vMQda9yYXJ-s,1560
 optimum/rbln/transformers/models/phi/__init__.py,sha256=mZLt1M7BbYEvSon5UlkniMUPa15SfjZFdw0kMSAF3VA,644
 optimum/rbln/transformers/models/phi/modeling_phi.py,sha256=j-6Pqd5rR2JE8I1pnKFlCi4nW5Dv3wZjoPWxohissoo,1516
-optimum/rbln/transformers/models/phi/phi_architecture.py,sha256=rBQjr6MOYBo1i5yLekMSR81TzYlHrHAA30kyKDdR7ww,4132
+optimum/rbln/transformers/models/phi/phi_architecture.py,sha256=TueyqmjPXWmOPOxBm4dIFyd0X3iV1jgw0U6c26iCAPk,4090
 optimum/rbln/transformers/models/qwen2/__init__.py,sha256=RAMWc21W_2I6DH9xBjeNxPECmAcTrbKhSIefq3Lass0,648
 optimum/rbln/transformers/models/qwen2/modeling_qwen2.py,sha256=9-aFDvjMzPNUyGOz0qo33RE18bUFGYZ3Wt_68zb5uJY,1530
 optimum/rbln/transformers/models/qwen2/qwen2_architecture.py,sha256=XlNAMYAcDLohnSAhIFGKOPuCB5XLgzYs5ABWdeQSaZs,720
@@ -114,7 +116,7 @@ optimum/rbln/utils/model_utils.py,sha256=DfD_Z2qvZHqcddXqnzTM1AN8khanj3-DXK2lJvV
 optimum/rbln/utils/runtime_utils.py,sha256=5-DYniyP59nx-mrrbi7AqA77L85b4Cm5oLpaxidSyss,3699
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=oZoGrItB8WqY4i-K9WJPlLlcLohc1YGB9OHB8_XZw3A,4071
-optimum_rbln-0.7.3a4.dist-info/METADATA,sha256=8VNTOVgsgFtcFUuZ9VEeRQfC2LEB60OFmW92hlJo8V8,5300
-optimum_rbln-0.7.3a4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.7.3a4.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.7.3a4.dist-info/RECORD,,
+optimum_rbln-0.7.3a6.dist-info/METADATA,sha256=TGw8TCIfBQ9RWlzxf5JI16Zoy-xoEodnBO8m6SKXBsk,5300
+optimum_rbln-0.7.3a6.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.7.3a6.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.7.3a6.dist-info/RECORD,,