0001-feat-enable-xpu-support-for-meta-reference-stack.patch

From 2018c3c386db716259b94b8ddcbea17e207b97a2 Mon Sep 17 00:00:00 2001
From: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Date: Mon, 18 Nov 2024 16:00:55 -0800
Subject: [PATCH] feat: enable xpu support for meta-reference stack

Requires: https://github.com/meta-llama/llama-models/pull/233
Signed-off-by: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
---
 .../inference/meta_reference/generation.py    | 45 +++++++++++++------
 1 file changed, 32 insertions(+), 13 deletions(-)

diff --git a/llama_stack/providers/inline/inference/meta_reference/generation.py b/llama_stack/providers/inline/inference/meta_reference/generation.py
index a96409ca..66f97b42 100644
--- a/llama_stack/providers/inline/inference/meta_reference/generation.py
+++ b/llama_stack/providers/inline/inference/meta_reference/generation.py
@@ -98,7 +98,10 @@ class Llama:
         """
         llama_model_id = llama_model.core_model_id.value
         if not torch.distributed.is_initialized():
-            torch.distributed.init_process_group("nccl")
+            if torch.cuda.is_available():
+                torch.distributed.init_process_group("nccl")
+            else:
+                torch.distributed.init_process_group("gloo")
 
         model_parallel_size = llama_model.pth_file_count
 
@@ -106,7 +109,14 @@ class Llama:
             initialize_model_parallel(model_parallel_size)
 
         local_rank = int(os.environ.get("LOCAL_RANK", 0))
-        torch.cuda.set_device(local_rank)
+        if torch.cuda.is_available():
+            device = "cuda"
+            torch.cuda.set_device(local_rank)
+        elif torch.xpu.is_available():
+            device = "xpu"
+            torch.xpu.set_device(local_rank)
+        else:
+            raise NotImplementedError("Devices other than CUDA and XPU are not supported yet")
 
         # seed must be the same in all processes
         if config.torch_seed is not None:
@@ -189,10 +199,17 @@ class Llama:
                     "Currently int4 and fp8 are the only supported quantization methods."
                 )
         else:
-            if torch.cuda.is_bf16_supported():
-                torch.set_default_tensor_type(torch.cuda.BFloat16Tensor)
-            else:
-                torch.set_default_tensor_type(torch.cuda.HalfTensor)
+            if device == "cuda":
+                if torch.cuda.is_bf16_supported():
+                    torch.set_default_tensor_type(torch.cuda.BFloat16Tensor)
+                else:
+                    torch.set_default_tensor_type(torch.cuda.HalfTensor)
+            elif device == "xpu":
+                torch.set_default_device(device)
+                if torch.xpu.is_bf16_supported():
+                    torch.set_default_dtype(torch.bfloat16)
+                else:
+                    torch.set_default_dtype(torch.half)
             if model_args.vision_chunk_size > 0:
                 model = CrossAttentionTransformer(model_args)
                 model.setup_cache(model_args.max_batch_size, torch.bfloat16)
@@ -200,6 +217,8 @@ class Llama:
                 model = Transformer(model_args)
             model.load_state_dict(state_dict, strict=False)
 
+        model.to(device)
+
         log.info(f"Loaded in {time.time() - start_time:.2f} seconds")
         return Llama(model, tokenizer, model_args, llama_model_id)
 
@@ -208,7 +227,7 @@ class Llama:
         model: Transformer,
         tokenizer: Tokenizer,
         args: ModelArgs,
-        llama_model: str,
+        llama_model: str
     ):
         self.args = args
         self.model = model
@@ -266,14 +285,14 @@ class Llama:
             )
 
         pad_id = self.tokenizer.pad_id
-        tokens = torch.full((bsz, total_len), pad_id, dtype=torch.long, device="cuda")
+        tokens = torch.full((bsz, total_len), pad_id, dtype=torch.long)
         for k, t in enumerate(prompt_tokens):
-            tokens[k, : len(t)] = torch.tensor(t, dtype=torch.long, device="cuda")
+            tokens[k, : len(t)] = torch.tensor(t, dtype=torch.long)
         if logprobs:
-            token_logprobs = torch.zeros_like(tokens, dtype=torch.float)
+            token_logprobs = torch.zeros_like(tokens)
 
         prev_pos = 0
-        eos_reached = torch.tensor([False] * bsz, device="cuda")
+        eos_reached = torch.tensor([False] * bsz)
         input_text_mask = tokens != pad_id
         if min_prompt_len == total_len:
             # TODO(ashwin): unify this branch with the one below and figure out multimodal crap
@@ -285,11 +304,11 @@ class Llama:
                 ignore_index=pad_id,
             )
 
-        stop_tokens = torch.tensor(self.tokenizer.stop_tokens, device="cuda")
+        stop_tokens = torch.tensor(self.tokenizer.stop_tokens)
         for cur_pos in range(min_prompt_len, total_len):
             if is_vision:
                 position_ids = torch.arange(
-                    prev_pos, cur_pos, dtype=torch.long, device="cuda"
+                    prev_pos, cur_pos, dtype=torch.long
                 )
                 logits = self.model.forward(
                     position_ids,
-- 
2.34.1