Update precision check for vicuna (#1610)

* set task_topology_max_group to cpu_count

by default. Can be overriden with a flag of the same str

* add download for int4/int8 mlir

apps/language_models/src/pipelines/vicuna_pipeline.py

@@ -38,8 +38,10 @@ class Vicuna(SharkLLMBase):
         super().__init__(model_name, hf_model_path, max_num_tokens)
         self.max_sequence_length = 256
         self.device = device
-        if precision in ["int4", "int8"]:
-            print("int4 and int8 are not supported yet, using fp32")
+        if not load_mlir_from_shark_tank and precision in ["int4", "int8"]:
+            print(
+                "int4 and int8 are only available from SHARK tank, please set --load_mlir_from_shark_tank, using fp32 now"
+            )
             precision = "fp32"
         self.precision = precision
         self.first_vicuna_vmfb_path = first_vicuna_vmfb_path
@@ -103,8 +105,8 @@ class Vicuna(SharkLLMBase):
         else:
             mlir_generated = False
             if self.load_mlir_from_shark_tank:
-                if self.precision in ["fp32", "fp16"]:
-                    # download MLIR from shark_tank for fp32/fp16
+                if self.precision in ["fp32", "fp16", "int8", "int4"]:
+                    # download MLIR from shark_tank
                     download_public_file(
                         f"gs://shark_tank/vicuna/unsharded/mlir/{self.first_vicuna_mlir_path.name}",
                         self.first_vicuna_mlir_path.absolute(),
@@ -121,7 +123,7 @@ class Vicuna(SharkLLMBase):
                         )
                 else:
                     print(
-                        f"Only fp32 and fp16 mlir added to tank, generating {self.precision} mlir on device."
+                        f"Only fp32/fp16/int8/int4 mlir added to tank, generating {self.precision} mlir on device."
                     )
 
             if not mlir_generated:
@@ -245,8 +247,8 @@ class Vicuna(SharkLLMBase):
         else:
             mlir_generated = False
             if self.load_mlir_from_shark_tank:
-                if self.precision in ["fp32", "fp16"]:
-                    # download MLIR from shark_tank for fp32/fp16
+                if self.precision in ["fp32", "fp16", "int8", "int4"]:
+                    # download MLIR from shark_tank
                     download_public_file(
                         f"gs://shark_tank/vicuna/unsharded/mlir/{self.second_vicuna_mlir_path.name}",
                         self.second_vicuna_mlir_path.absolute(),
@@ -263,7 +265,7 @@ class Vicuna(SharkLLMBase):
                         )
                 else:
                     print(
-                        "Only fp32 mlir added to tank, generating mlir on device."
+                        "Only fp32/fp16/int8/int4 mlir added to tank, generating mlir on device."
                     )
 
             if not mlir_generated:
@@ -439,6 +441,14 @@ class Vicuna(SharkLLMBase):
         # return tuple of shark_modules once mem is supported
         # return fvic_shark_model, svic_shark_model
 
+    def decode_tokens(self, res_tokens):
+        for i in range(len(res_tokens)):
+            if type(res_tokens[i]) != int:
+                res_tokens[i] = int(res_tokens[i][0])
+
+        res_str = self.tokenizer.decode(res_tokens)
+        return res_str
+
     def generate(self, prompt, cli=False):
         # TODO: refactor for cleaner integration
         import gc
@@ -448,7 +458,6 @@ class Vicuna(SharkLLMBase):
                 self.first_vic = self.compile_first_vicuna()
             if self.second_vic == None:
                 self.second_vic = self.compile_second_vicuna()
-        res = []
         res_tokens = []
         params = {
             "prompt": prompt,
@@ -464,8 +473,8 @@ class Vicuna(SharkLLMBase):
         logits = generated_token_op["logits"]
         pkv = generated_token_op["pkv"]
         detok = generated_token_op["detok"]
+        yield detok
 
-        res.append(detok)
         res_tokens.append(token)
         if cli:
             print(f"Assistant: {detok}", end=" ", flush=True)
@@ -498,25 +507,24 @@ class Vicuna(SharkLLMBase):
                 break
             res_tokens.append(token)
             if detok == "<0x0A>":
-                res.append("\n")
                 if cli:
                     print("\n", end="", flush=True)
             else:
-                res.append(detok)
                 if cli:
                     print(f"{detok}", end=" ", flush=True)
+
+            if len(res_tokens) % 3 == 0:
+                part_str = self.decode_tokens(res_tokens)
+                yield part_str
+
         if self.device == "cuda":
             del sec_vic, pkv, logits
             torch.cuda.empty_cache()
             gc.collect()
 
-        for i in range(len(res_tokens)):
-            if type(res_tokens[i]) != int:
-                res_tokens[i] = int(res_tokens[i][0])
-
-        res_str = self.tokenizer.decode(res_tokens)
+        res_str = self.decode_tokens(res_tokens)
         # print(f"[DEBUG] final output : \n{res_str}")
-        return res_str
+        yield res_str
 
     def generate_new_token(self, params, debug=False):
         def forward_first(first_vic, prompt, cache_outputs=False):

apps/stable_diffusion/scripts/img2img.py

@@ -103,6 +103,7 @@ def main():
         dtype,
         args.use_base_vae,
         cpu_scheduling,
+        args.max_embeddings_multiples,
         use_stencil=use_stencil,
     )
     total_time = time.time() - start_time

apps/stable_diffusion/scripts/inpaint.py

@@ -81,6 +81,7 @@ def main():
             dtype,
             args.use_base_vae,
             cpu_scheduling,
+            args.max_embeddings_multiples,
         )
         total_time = time.time() - start_time
         text_output = f"prompt={args.prompts}"

apps/stable_diffusion/scripts/outpaint.py

@@ -79,6 +79,7 @@ def main():
             dtype,
             args.use_base_vae,
             cpu_scheduling,
+            args.max_embeddings_multiples,
         )
         total_time = time.time() - start_time
         text_output = f"prompt={args.prompts}"

apps/stable_diffusion/scripts/upscaler.py

@@ -73,6 +73,7 @@ if __name__ == "__main__":
         dtype,
         args.use_base_vae,
         cpu_scheduling,
+        args.max_embeddings_multiples,
     )
     total_time = time.time() - start_time
     text_output = f"prompt={args.prompts}"

apps/stable_diffusion/shark_sd.spec

@@ -78,7 +78,7 @@ exe = EXE(
     debug=False,
     bootloader_ignore_signals=False,
     strip=False,
-    upx=True,
+    upx=False,
     upx_exclude=[],
     runtime_tmpdir=None,
     console=True,

apps/stable_diffusion/src/models/model_wrappers.py

@@ -520,16 +520,17 @@ class SharkifyStableDiffusionModel:
                 torch.nn.functional.pad(inputs[2], pad),
                 inputs[3])
         input_mask = [True, True, True, False]
+        model_name = "unet512" if use_large else "unet"
         shark_unet, unet_mlir = compile_through_fx(
             unet,
             inputs,
-            extended_model_name=self.model_name["unet"],
+            extended_model_name=self.model_name[model_name],
             is_f16=is_f16,
             f16_input_mask=input_mask,
             use_tuned=self.use_tuned,
             extra_args=get_opt_flags("unet", precision=self.precision),
             base_model_id=self.base_model_id,
-            model_name="unet",
+            model_name=model_name,
             precision=self.precision,
             return_mlir=self.return_mlir,
         )

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_img2img.py

@@ -135,6 +135,7 @@ class Image2ImagePipeline(StableDiffusionPipeline):
         dtype,
         use_base_vae,
         cpu_scheduling,
+        max_embeddings_multiples,
         use_stencil,
     ):
         # prompts and negative prompts must be a list.
@@ -156,7 +157,10 @@ class Image2ImagePipeline(StableDiffusionPipeline):
 
         # Get text embeddings with weight emphasis from prompts
         text_embeddings = self.encode_prompts_weight(
-            prompts, neg_prompts, max_length
+            prompts,
+            neg_prompts,
+            max_length,
+            max_embeddings_multiples=max_embeddings_multiples,
         )
 
         # guidance scale as a float32 tensor.

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_inpaint.py

@@ -378,6 +378,7 @@ class InpaintPipeline(StableDiffusionPipeline):
         dtype,
         use_base_vae,
         cpu_scheduling,
+        max_embeddings_multiples,
     ):
         # prompts and negative prompts must be a list.
         if isinstance(prompts, str):
@@ -408,7 +409,10 @@ class InpaintPipeline(StableDiffusionPipeline):
 
         # Get text embeddings with weight emphasis from prompts
         text_embeddings = self.encode_prompts_weight(
-            prompts, neg_prompts, max_length
+            prompts,
+            neg_prompts,
+            max_length,
+            max_embeddings_multiples=max_embeddings_multiples,
         )
 
         # guidance scale as a float32 tensor.

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_outpaint.py

@@ -379,6 +379,7 @@ class OutpaintPipeline(StableDiffusionPipeline):
         dtype,
         use_base_vae,
         cpu_scheduling,
+        max_embeddings_multiples,
     ):
         # prompts and negative prompts must be a list.
         if isinstance(prompts, str):
@@ -409,7 +410,10 @@ class OutpaintPipeline(StableDiffusionPipeline):
 
         # Get text embeddings with weight emphasis from prompts
         text_embeddings = self.encode_prompts_weight(
-            prompts, neg_prompts, max_length
+            prompts,
+            neg_prompts,
+            max_length,
+            max_embeddings_multiples=max_embeddings_multiples,
         )
 
         # guidance scale as a float32 tensor.

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_stencil.py

@@ -204,6 +204,7 @@ class StencilPipeline(StableDiffusionPipeline):
         dtype,
         use_base_vae,
         cpu_scheduling,
+        max_embeddings_multiples,
         use_stencil,
     ):
         # Control Embedding check & conversion
@@ -230,7 +231,10 @@ class StencilPipeline(StableDiffusionPipeline):
 
         # Get text embeddings with weight emphasis from prompts
         text_embeddings = self.encode_prompts_weight(
-            prompts, neg_prompts, max_length
+            prompts,
+            neg_prompts,
+            max_length,
+            max_embeddings_multiples=max_embeddings_multiples,
         )
 
         # guidance scale as a float32 tensor.

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_upscaler.py

@@ -168,7 +168,10 @@ class UpscalerPipeline(StableDiffusionPipeline):
         text_embeddings = torch.from_numpy(text_embeddings).to(dtype)
         text_embeddings_numpy = text_embeddings.detach().numpy()
         self.status = SD_STATE_IDLE
-        self.load_unet()
+        if text_embeddings.shape[1] <= self.model_max_length:
+            self.load_unet()
+        else:
+            self.load_unet_512()
         for i, t in tqdm(enumerate(total_timesteps)):
             step_start_time = time.time()
             latent_model_input = torch.cat([latents] * 2)
@@ -182,15 +185,26 @@ class UpscalerPipeline(StableDiffusionPipeline):
 
             # Profiling Unet.
             profile_device = start_profiling(file_path="unet.rdc")
-            noise_pred = self.unet(
-                "forward",
-                (
-                    latent_model_input,
-                    timestep,
-                    text_embeddings_numpy,
-                    noise_level,
-                ),
-            )
+            if text_embeddings.shape[1] <= self.model_max_length:
+                noise_pred = self.unet(
+                    "forward",
+                    (
+                        latent_model_input,
+                        timestep,
+                        text_embeddings_numpy,
+                        noise_level,
+                    ),
+                )
+            else:
+                noise_pred = self.unet_512(
+                    "forward",
+                    (
+                        latent_model_input,
+                        timestep,
+                        text_embeddings_numpy,
+                        noise_level,
+                    ),
+                )
             end_profiling(profile_device)
             noise_pred = torch.from_numpy(noise_pred)
             noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
@@ -219,6 +233,7 @@ class UpscalerPipeline(StableDiffusionPipeline):
 
         if self.ondemand:
             self.unload_unet()
+            self.unload_unet_512()
         avg_step_time = step_time_sum / len(total_timesteps)
         self.log += f"\nAverage step time: {avg_step_time}ms/it"
 
@@ -243,6 +258,7 @@ class UpscalerPipeline(StableDiffusionPipeline):
         dtype,
         use_base_vae,
         cpu_scheduling,
+        max_embeddings_multiples,
     ):
         # prompts and negative prompts must be a list.
         if isinstance(prompts, str):
@@ -264,7 +280,10 @@ class UpscalerPipeline(StableDiffusionPipeline):
 
         # Get text embeddings with weight emphasis from prompts
         text_embeddings = self.encode_prompts_weight(
-            prompts, neg_prompts, max_length
+            prompts,
+            neg_prompts,
+            max_length,
+            max_embeddings_multiples=max_embeddings_multiples,
         )
 
         # 4. Preprocess image

apps/stable_diffusion/src/utils/utils.py

@@ -810,8 +810,11 @@ def save_output_img(output_img, img_seed, extra_info={}):
 
     new_entry.update(extra_info)
 
-    with open(csv_path, "a", encoding="utf-8") as csv_obj:
+    csv_mode = "a" if os.path.isfile(csv_path) else "w"
+    with open(csv_path, csv_mode, encoding="utf-8") as csv_obj:
         dictwriter_obj = DictWriter(csv_obj, fieldnames=list(new_entry.keys()))
+        if csv_mode == "w":
+            dictwriter_obj.writeheader()
         dictwriter_obj.writerow(new_entry)
         csv_obj.close()
 

apps/stable_diffusion/web/index.py

@@ -30,7 +30,11 @@ def launch_app(address):
     width = window.winfo_screenwidth()
     height = window.winfo_screenheight()
     webview.create_window(
-        "SHARK AI Studio", url=address, width=width, height=height
+        "SHARK AI Studio",
+        url=address,
+        width=width,
+        height=height,
+        text_select=True,
     )
     webview.start(private_mode=False)
 

apps/stable_diffusion/web/ui/img2img_ui.py

@@ -249,6 +249,7 @@ def img2img_inf(
             dtype,
             args.use_base_vae,
             cpu_scheduling,
+            args.max_embeddings_multiples,
             use_stencil=use_stencil,
         )
         seeds.append(img_seed)

apps/stable_diffusion/web/ui/inpaint_ui.py

@@ -204,6 +204,7 @@ def inpaint_inf(
             dtype,
             args.use_base_vae,
             cpu_scheduling,
+            args.max_embeddings_multiples,
         )
         seeds.append(img_seed)
         total_time = time.time() - start_time

apps/stable_diffusion/web/ui/outpaint_ui.py

@@ -211,6 +211,7 @@ def outpaint_inf(
             dtype,
             args.use_base_vae,
             cpu_scheduling,
+            args.max_embeddings_multiples,
         )
         seeds.append(img_seed)
         total_time = time.time() - start_time

apps/stable_diffusion/web/ui/stablelm_ui.py

@@ -65,16 +65,11 @@ def chat(curr_system_message, history, model, device, precision):
         )
         prompt = messages.strip()
         print("prompt = ", prompt)
-        sentence = vicuna_model.generate(prompt)
 
-        partial_text = ""
-        for new_text in sentence.split(" "):
-            # print(new_text)
-            partial_text += new_text + " "
+        for partial_text in vicuna_model.generate(prompt):
             history[-1][1] = partial_text
-            # Yield an empty string to cleanup the message textbox and the updated conversation history
             yield history
-        history[-1][1] = sentence
+
         return history
 
     # else Model is StableLM

apps/stable_diffusion/web/ui/upscaler_ui.py

@@ -202,6 +202,7 @@ def upscaler_inf(
                     dtype,
                     args.use_base_vae,
                     cpu_scheduling,
+                    args.max_embeddings_multiples,
                 )
                 if global_obj.get_sd_status() == SD_STATE_CANCEL:
                     break

apps/stable_diffusion/web/utils/metadata/csv_metadata.py

@@ -11,21 +11,35 @@ def has_csv(image_filename: str) -> bool:
     return os.path.exists(csv_path(image_filename))
 
 
-def parse_csv(image_filename: str):
-    # We use a reader instead of a DictReader here for images_details.csv files due to the lack of
-    # headers, and then match up the return list for each row with our guess at which column format
-    # the file is using.
-
+def matching_filename(image_filename: str, row):
     # we assume the final column of the csv has the original filename with full path and match that
-    # against the image_filename. We then exclude the filename from the output, hence the -1's.
+    # against the image_filename if we are given a list. Otherwise we assume a dict and and take
+    # the value of the OUTPUT key
+    return os.path.basename(image_filename) in (
+        row[-1] if isinstance(row, list) else row["OUTPUT"]
+    )
+
+
+def parse_csv(image_filename: str):
     csv_filename = csv_path(image_filename)
 
-    matches = [
-        humanize(row)
-        for row in csv.reader(open(csv_filename, "r", newline=""))
-        if row
-        and humanizable(row)
-        and os.path.basename(image_filename) in row[-1]
-    ]
+    with open(csv_filename, "r", newline="") as csv_file:
+        # We use a reader or DictReader here for images_details.csv depending on whether we think it
+        # has headers or not. Having headers means less guessing of the format.
+        has_header = csv.Sniffer().has_header(csv_file.read(2048))
+        csv_file.seek(0)
+
+        reader = (
+            csv.DictReader(csv_file) if has_header else csv.reader(csv_file)
+        )
+
+        matches = [
+            # we rely on humanize and humanizable to work out the parsing of the individual .csv rows
+            humanize(row)
+            for row in reader
+            if row
+            and (has_header or humanizable(row))
+            and matching_filename(image_filename, row)
+        ]
 
     return matches[0] if matches else {}

apps/stable_diffusion/web/utils/metadata/format.py

@@ -50,7 +50,22 @@ PARAMS_FORMATS = {
     },
 }
 
-PARAMS_FORMAT_LONGEST = PARAMS_FORMATS[max(PARAMS_FORMATS.keys())]
+PARAMS_FORMAT_CURRENT = {
+    "VARIANT": "Model",
+    "VAE": "VAE",
+    "LORA": "LoRA",
+    "SCHEDULER": "Sampler",
+    "PROMPT": "Prompt",
+    "NEG_PROMPT": "Negative prompt",
+    "SEED": "Seed",
+    "CFG_SCALE": "CFG scale",
+    "PRECISION": "Precision",
+    "STEPS": "Steps",
+    "HEIGHT": "Height",
+    "WIDTH": "Width",
+    "MAX_LENGTH": "Max Length",
+    "OUTPUT": "Filename",
+}
 
 
 def compact(metadata: dict) -> dict:
@@ -97,19 +112,20 @@ def humanize(metadata: dict | list[str], includes_filename=True) -> dict:
             )
 
     # For dictionaries we try to use the matching length parameter format if
-    # available, otherwise we use the longest. Then we swap keys in the
-    # metadata that match keys in the format for the friendlier name that we
-    # have set in the format value
+    # available, otherwise we just use the current format which is assumed to
+    # have everything currently known about. Then we swap keys in the metadata
+    # that match keys in the format for the friendlier name that we have set
+    # in the format value
     if isinstance(metadata, dict):
         if humanizable(metadata, includes_filename):
             format = PARAMS_FORMATS[lookup_key]
         else:
-            format = PARAMS_FORMAT_LONGEST
+            format = PARAMS_FORMAT_CURRENT
 
         return {
-            format[key]: value
-            for (key, value) in metadata.items()
-            if key in format.keys()
+            format[key]: metadata[key]
+            for key in format.keys()
+            if key in metadata.keys() and metadata[key]
         }
 
     raise TypeError("Can only humanize parameter lists or dictionaries")

shark/dynamo_backend/utils.py

@@ -0,0 +1,154 @@
+import functools
+from typing import List, Optional
+import torch
+from torch.fx.experimental.proxy_tensor import make_fx
+from torch._functorch.compile_utils import strip_overloads
+from shark.shark_inference import SharkInference
+from torch._decomp import get_decompositions
+from torch.func import functionalize
+import io
+import torch_mlir
+
+
+# TODO: Control decompositions.
+def default_decompositions():
+    return get_decompositions(
+        [
+            torch.ops.aten.embedding_dense_backward,
+            torch.ops.aten.native_layer_norm_backward,
+            torch.ops.aten.slice_backward,
+            torch.ops.aten.select_backward,
+            torch.ops.aten.norm.ScalarOpt_dim,
+            torch.ops.aten.native_group_norm,
+            torch.ops.aten.upsample_bilinear2d.vec,
+            torch.ops.aten.split.Tensor,
+            torch.ops.aten.split_with_sizes,
+            torch.ops.aten.native_layer_norm,
+            torch.ops.aten.masked_fill.Tensor,
+            torch.ops.aten.masked_fill.Scalar,
+        ]
+    )
+
+
+def _remove_nones(fx_g: torch.fx.GraphModule) -> List[int]:
+    removed_indexes = []
+    for node in fx_g.graph.nodes:
+        if node.op == "output":
+            assert (
+                len(node.args) == 1
+            ), "Output node must have a single argument"
+            node_arg = node.args[0]
+            if isinstance(node_arg, (list, tuple)):
+                node_arg = list(node_arg)
+                node_args_len = len(node_arg)
+                for i in range(node_args_len):
+                    curr_index = node_args_len - (i + 1)
+                    if node_arg[curr_index] is None:
+                        removed_indexes.append(curr_index)
+                        node_arg.pop(curr_index)
+                node.args = (tuple(node_arg),)
+                break
+
+    if len(removed_indexes) > 0:
+        fx_g.graph.lint()
+        fx_g.graph.eliminate_dead_code()
+        fx_g.recompile()
+    removed_indexes.sort()
+    return removed_indexes
+
+
+def _returns_nothing(fx_g: torch.fx.GraphModule) -> bool:
+    for node in fx_g.graph.nodes:
+        if node.op == "output":
+            assert (
+                len(node.args) == 1
+            ), "Output node must have a single argument"
+            node_arg = node.args[0]
+            if isinstance(node_arg, tuple):
+                return len(node_arg) == 0
+    return False
+
+
+def _unwrap_single_tuple_return(fx_g: torch.fx.GraphModule) -> bool:
+    """
+    Replace tuple with tuple element in functions that return one-element tuples.
+    Returns true if an unwrapping took place, and false otherwise.
+    """
+    unwrapped_tuple = False
+    for node in fx_g.graph.nodes:
+        if node.op == "output":
+            assert (
+                len(node.args) == 1
+            ), "Output node must have a single argument"
+            node_arg = node.args[0]
+            if isinstance(node_arg, tuple):
+                if len(node_arg) == 1:
+                    node.args = (node_arg[0],)
+                    unwrapped_tuple = True
+                    break
+
+    if unwrapped_tuple:
+        fx_g.graph.lint()
+        fx_g.recompile()
+    return unwrapped_tuple
+
+
+class SharkBackend:
+    def __init__(
+        self, fx_g: torch.fx.GraphModule, inputs: tuple, options: dict
+    ):
+        self.fx_g = fx_g
+        self.inputs = inputs
+        self.shark_module = None
+        self.device: str = options.get("device", "cpu")
+        self.was_unwrapped: bool = False
+        self.none_indices: list = []
+        self._modify_fx_g()
+        self.compile()
+
+    def _modify_fx_g(self):
+        self.none_indices = _remove_nones(self.fx_g)
+        self.was_unwrapped = _unwrap_single_tuple_return(self.fx_g)
+
+    def compile(self):
+        gm = make_fx(
+            functionalize(self.fx_g),
+            decomposition_table=default_decompositions(),
+        )(*self.inputs)
+        gm.graph.set_codegen(torch.fx.graph.CodeGen())
+        gm.recompile()
+        strip_overloads(gm)
+        ts_g = torch.jit.script(gm)
+        mlir_module = torch_mlir.compile(
+            ts_g, self.inputs, output_type="linalg-on-tensors"
+        )
+        bytecode_stream = io.BytesIO()
+        mlir_module.operation.write_bytecode(bytecode_stream)
+        bytecode = bytecode_stream.getvalue()
+        from shark.shark_inference import SharkInference
+
+        shark_module = SharkInference(
+            mlir_module=bytecode,
+            device=self.device,
+            mlir_dialect="tm_tensor",
+        )
+        shark_module.compile(extra_args=[])
+        self.shark_module = shark_module
+
+    def __call__(self, *inputs):
+        np_inputs = [x.contiguous().detach().cpu().numpy() for x in inputs]
+        np_outs = self.shark_module("forward", np_inputs)
+        if self.was_unwrapped:
+            np_outs = [
+                np_outs,
+            ]
+
+        if not isinstance(np_outs, list):
+            res = torch.from_numpy(np_outs)
+            return res
+
+        result = [torch.from_numpy(x) for x in np_outs]
+        for r_in in self.none_indices:
+            result.insert(r_in, None)
+        result = tuple(result)
+        return result

shark/examples/shark_inference/mega_test.py

@@ -1,10 +1,7 @@
 import torch
 import torch_mlir
 from shark.shark_inference import SharkInference
-from apps.stable_diffusion.src.utils import (
-    compile_through_fx,
-    args,
-)
+from shark.shark_compile import shark_compile_through_fx
 from MEGABYTE_pytorch import MEGABYTE
 
 import os
@@ -37,23 +34,22 @@ class MegaModel(torch.nn.Module):
 
 
 megaModel = MegaModel()
-input = [torch.randint(0, 16000, (1, 1024, 4))]
+inputs = [torch.randint(0, 16000, (1, 1024, 4))]
 
 # CURRENTLY IT BAILS OUT HERE BECAUSE OF MISSING OP LOWERINGS :-
 # 1. aten.alias
-shark_module, _ = compile_through_fx(
-    megaModel,
-    inputs=input,
+shark_module, _ = shark_compile_through_fx(
+    model=megaModel,
+    inputs=inputs,
     extended_model_name="mega_shark",
-    debug=False,
-    generate_vmfb=True,
+    is_f16=False,
+    f16_input_mask=None,
     save_dir=os.getcwd(),
+    debug=False,
+    generate_or_load_vmfb=True,
     extra_args=[],
-    base_model_id=None,
-    model_name="mega_shark",
-    precision=None,
-    return_mlir=True,
     device="cuda",
+    mlir_dialect="tm_tensor",
 )
 # logits = model(x)
 
@@ -63,10 +59,10 @@ def print_output_info(output, msg):
     print("\n\t", output.shape)
 
 
-ans = shark_module("forward", input)
+ans = shark_module("forward", inputs)
 print_output_info(torch.from_numpy(ans), "SHARK's output")
 
-ans = megaModel.forward(*input)
+ans = megaModel.forward(*inputs)
 print_output_info(ans, "ORIGINAL Model's output")
 
 # and sample from the logits accordingly

shark/iree_utils/compile_utils.py

@@ -14,6 +14,7 @@
 import iree.runtime as ireert
 import iree.compiler as ireec
 from shark.iree_utils._common import iree_device_map, iree_target_map
+from shark.iree_utils.cpu_utils import get_iree_cpu_rt_args
 from shark.iree_utils.benchmark_utils import *
 from shark.parser import shark_args
 import numpy as np
@@ -352,6 +353,12 @@ def load_vmfb_using_mmap(
         config = ireert.Config(device=haldevice)
     else:
         config = get_iree_runtime_config(device)
+    if "task" in device:
+        print(
+            f"[DEBUG] setting iree runtime flags for cpu:\n{' '.join(get_iree_cpu_rt_args())}"
+        )
+        for flag in get_iree_cpu_rt_args():
+            ireert.flags.parse_flags(flag)
     # Now load vmfb.
     # Two scenarios we have here :-
     #      1. We either have the vmfb already saved and therefore pass the path of it.
@@ -359,7 +366,6 @@ def load_vmfb_using_mmap(
     #   OR 2. We are compiling on the fly, therefore we have the flatbuffer blob to play with.
     #         (This would arise if we're invoking `compile` from a SharkInference obj)
     temp_file_to_unlink = None
-
     if isinstance(flatbuffer_blob_or_path, Path):
         flatbuffer_blob_or_path = flatbuffer_blob_or_path.__str__()
     if (

shark/iree_utils/cpu_utils.py

@@ -16,6 +16,7 @@
 
 import subprocess
 import platform
+from shark.parser import shark_args
 
 
 def get_cpu_count():
@@ -44,4 +45,18 @@ def get_iree_cpu_args():
         error_message = f"OS Type f{os_name} not supported and triple can't be determined, open issue to dSHARK team please :)"
         raise Exception(error_message)
     print(f"Target triple found:{target_triple}")
-    return [f"--iree-llvmcpu-target-triple={target_triple}"]
+    return [
+        f"--iree-llvmcpu-target-triple={target_triple}",
+    ]
+
+
+# Get iree runtime flags for cpu
+def get_iree_cpu_rt_args():
+    default = get_cpu_count()
+    default = default if default <= 8 else default - 2
+    cpu_count = (
+        default
+        if shark_args.task_topology_max_group_count is None
+        else shark_args.task_topology_max_group_count
+    )
+    return [f"--task_topology_max_group_count={cpu_count}"]

shark/parser.py

@@ -119,5 +119,11 @@ parser.add_argument(
     "to augment the base device allocator",
     choices=["debug", "caching"],
 )
+parser.add_argument(
+    "--task_topology_max_group_count",
+    type=str,
+    default=None,
+    help="passthrough flag for the iree flag of the same name. If None, defaults to cpu-count",
+)
 
 shark_args, unknown = parser.parse_known_args()

shark/shark_compile.py

@@ -0,0 +1,99 @@
+import os
+import tempfile
+from shark.shark_inference import SharkInference
+from shark.shark_importer import import_with_fx
+
+
+def load_vmfb(extended_model_name, device, mlir_dialect, extra_args=[]):
+    vmfb_path = os.path.join(os.getcwd(), extended_model_name + ".vmfb")
+    shark_module = None
+    if os.path.isfile(vmfb_path):
+        shark_module = SharkInference(
+            None,
+            device=device,
+            mlir_dialect=mlir_dialect,
+        )
+        print(f"loading existing vmfb from: {vmfb_path}")
+        shark_module.load_module(vmfb_path, extra_args=extra_args)
+    return shark_module
+
+
+def compile_module(
+    shark_module, extended_model_name, generate_vmfb, extra_args=[]
+):
+    if generate_vmfb:
+        vmfb_path = os.path.join(os.getcwd(), extended_model_name + ".vmfb")
+        if os.path.isfile(vmfb_path):
+            print(f"loading existing vmfb from: {vmfb_path}")
+            shark_module.load_module(vmfb_path, extra_args=extra_args)
+        else:
+            print(
+                "No vmfb found. Compiling and saving to {}".format(vmfb_path)
+            )
+            path = shark_module.save_module(
+                os.getcwd(), extended_model_name, extra_args
+            )
+            shark_module.load_module(path, extra_args=extra_args)
+    else:
+        shark_module.compile(extra_args)
+    return shark_module
+
+
+def shark_compile_through_fx(
+    model,
+    inputs,
+    extended_model_name,
+    is_f16=False,
+    f16_input_mask=None,
+    save_dir=tempfile.gettempdir(),
+    debug=False,
+    generate_or_load_vmfb=True,
+    extra_args=[],
+    device=None,
+    mlir_dialect="tm_tensor",
+):
+    if generate_or_load_vmfb:
+        shark_module = load_vmfb(
+            extended_model_name=extended_model_name,
+            device=device,
+            mlir_dialect=mlir_dialect,
+            extra_args=extra_args,
+        )
+        if shark_module:
+            return (
+                shark_module,
+                None,
+            )
+
+    from shark.parser import shark_args
+
+    if "cuda" in device:
+        shark_args.enable_tf32 = True
+
+    (
+        mlir_module,
+        _,
+    ) = import_with_fx(
+        model=model,
+        inputs=inputs,
+        is_f16=is_f16,
+        f16_input_mask=f16_input_mask,
+        debug=debug,
+        model_name=extended_model_name,
+        save_dir=save_dir,
+    )
+
+    shark_module = SharkInference(
+        mlir_module,
+        device=device,
+        mlir_dialect=mlir_dialect,
+    )
+    return (
+        compile_module(
+            shark_module,
+            extended_model_name,
+            generate_vmfb=generate_or_load_vmfb,
+            extra_args=extra_args,
+        ),
+        mlir_module,
+    )

shark/sharkdynamo/README.md

@@ -1,11 +0,0 @@
-1. Install torchdynamo
-   - `git clone https://github.com/pytorch/torchdynamo.git`
-   - `cd torchdynamo`
-   - `python -m pip install -r requirements.txt`
-   - `python setup.py develop`
-
-2. Install functorch
-   - `python -m pip install -v "git+https://github.com/pytorch/pytorch.git@$(python -c "import torch.version; print(torch.version.git_version)")#subdirectory=functorch"`
-
-3. Run examples.
-    - `python shark/examples/shark_dynamo/basic_examples.py`

shark/sharkdynamo/utils.py

@@ -1,163 +0,0 @@
-import functools
-import time
-from typing import List, Optional
-import torch
-from torch.fx.experimental.proxy_tensor import make_fx
-from torch._functorch.compile_utils import strip_overloads
-from shark.shark_inference import SharkInference
-from torch._decomp import get_decompositions
-
-import torch_mlir
-
-
-# TODO: Control decompositions.
-def default_decompositions():
-    return get_decompositions(
-        [
-            torch.ops.aten.embedding_dense_backward,
-            torch.ops.aten.native_layer_norm_backward,
-            torch.ops.aten.slice_backward,
-            torch.ops.aten.select_backward,
-            torch.ops.aten.norm.ScalarOpt_dim,
-            torch.ops.aten.native_group_norm,
-            torch.ops.aten.upsample_bilinear2d.vec,
-            torch.ops.aten.split.Tensor,
-            torch.ops.aten.split_with_sizes,
-        ]
-    )
-
-
-def timeit(*, append_time_to: Optional[List] = None):
-    def decorator(func):
-        @functools.wraps(func)
-        def wrapper(*args, **kwargs):
-            start_time = time.time_ns()
-            result = func(*args, **kwargs)
-            end_time = time.time_ns()
-
-            if append_time_to is not None:
-                append_time_to.append(end_time - start_time)
-            return result
-
-        return wrapper
-
-    return decorator
-
-
-def _returns_nothing(fx_g: torch.fx.GraphModule) -> bool:
-    for node in fx_g.graph.nodes:
-        if node.op == "output":
-            assert (
-                len(node.args) == 1
-            ), "Output node must have a single argument"
-            node_arg = node.args[0]
-            if isinstance(node_arg, tuple):
-                return len(node_arg) == 0
-    return False
-
-
-def _unwrap_single_tuple_return(fx_g: torch.fx.GraphModule) -> bool:
-    """
-    Replace tuple with tuple element in functions that return one-element tuples.
-    Returns true if an unwrapping took place, and false otherwise.
-    """
-    unwrapped_tuple = False
-    for node in fx_g.graph.nodes:
-        if node.op == "output":
-            assert (
-                len(node.args) == 1
-            ), "Output node must have a single argument"
-            node_arg = node.args[0]
-            if isinstance(node_arg, tuple):
-                if len(node_arg) == 1:
-                    node.args = (node_arg[0],)
-                    unwrapped_tuple = True
-                    break
-
-    if unwrapped_tuple:
-        fx_g.graph.lint()
-        fx_g.recompile()
-    return unwrapped_tuple
-
-
-def make_shark_compiler(use_tracing: bool, device: str, verbose=False):
-    def compiler(
-        fx_graph: torch.fx.GraphModule,
-        example_inputs: List[torch.Tensor],
-    ):
-        """Compile GraphModule using torch-mlir + SHARK."""
-        if verbose:
-            print("Compiling graph...")
-
-        if _returns_nothing(fx_graph):
-            return fx_graph
-
-        was_unwrapped = _unwrap_single_tuple_return(fx_graph)
-        fx_graph = make_fx(
-            fx_graph, decomposition_table=default_decompositions()
-        )(*example_inputs)
-        strip_overloads(fx_graph)
-
-        if verbose:
-            print("torch.fx graph:")
-            print(fx_graph.graph)
-
-        ts_compiler = torch.jit.trace if use_tracing else torch.jit.script
-        ts_graph = ts_compiler(fx_graph, example_inputs)
-
-        if verbose:
-            torch_mlir_module = torch_mlir.compile(
-                ts_graph,
-                example_inputs,
-                output_type=torch_mlir.OutputType.TORCH,
-            )
-            print("\n\ntorch-mlir backend contract graph:")
-            print(torch_mlir_module)
-
-        linalg_module = torch_mlir.compile(
-            ts_graph,
-            example_inputs,
-            output_type=torch_mlir.OutputType.LINALG_ON_TENSORS,
-        )
-        import io
-
-        bytecode_stream = io.BytesIO()
-        linalg_module.operation.write_bytecode(bytecode_stream)
-        mlir_module = bytecode_stream.getvalue()
-
-        shark_module = SharkInference(
-            mlir_module, mlir_dialect="linalg", device=device
-        )
-        shark_module.compile()
-
-        def forward(*inputs):
-            result = shark_module("forward", inputs)
-            result = tuple() if result is None else result
-            return (result,) if was_unwrapped else result
-
-        return forward
-
-    return compiler
-
-
-def check_results(compiled_results, eager_results):
-    for compiled_result, eager_result in zip(compiled_results, eager_results):
-        if not torch.allclose(
-            compiled_result.to("cpu"), eager_result.to("cpu"), atol=1e-5
-        ):
-            print("Compiled result does not match eager result")
-            return
-    print("Compiled result matches eager result!")
-
-
-def print_time_stats(times):
-    times_tensor = torch.tensor(times)
-
-    def quantile_ms(q):
-        return torch.quantile(times_tensor.to(float), q).item() / 1e6
-
-    print(f"Median: {quantile_ms(0.5)} ms")
-    print(f"10%ile: {quantile_ms(0.1)} ms")
-    print(f"90%ile: {quantile_ms(0.9)} ms")
-    print(f"Total: {torch.sum(times_tensor) / 1e6} ms")
-    print()