feat: enable xpu support for meta-reference stack (#558)

This commit adds support for XPU and CPU devices into meta-reference
stack for text models. On creation stack automatically identifies which
device to use checking available accelerate capabilities in the
following order: CUDA, then XPU, finally CPU. This behaviour can be
overwritten with the `DEVICE` environment variable. In this case
explicitly specified device will be used.

Tested with:
```
torchrun pytest llama_stack/providers/tests/inference/test_text_inference.py -k meta_reference
```

Results:
* Tested on: system with single CUDA device, system with single XPU
device and on pure CPU system
* Results: all test pass except `test_completion_logprobs`
* `test_completion_logprobs` fails in the same way as on a baseline,
i.e. unrelated with this change: `AssertionError: Unexpected top_k=3`

Requires: https://github.com/meta-llama/llama-models/pull/233

Signed-off-by: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>

llama_stack/providers/inline/inference/meta_reference/generation.py

@@ -96,9 +96,27 @@ class Llama:
             This method initializes the distributed process group, sets the device to CUDA,
             and loads the pre-trained model and tokenizer.
         """
+        if "DEVICE" in os.environ:
+            device = os.environ.get("DEVICE")
+            if device == "cuda":
+                assert torch.cuda.is_available(), "PyTorch CUDA backend not available"
+            if device == "xpu":
+                assert torch.xpu.is_available(), "PyTorch XPU backend not available"
+        else:
+            if torch.cuda.is_available():
+                device = "cuda"
+            elif torch.xpu.is_available():
+                device = "xpu"
+            else:
+                device = "cpu"
+        log.info(f"Using {device} device")
+
         llama_model_id = llama_model.core_model_id.value
         if not torch.distributed.is_initialized():
+            if device == "cuda":
                 torch.distributed.init_process_group("nccl")
+            else:
+                torch.distributed.init_process_group("gloo")
 
         model_parallel_size = llama_model.pth_file_count
 
@@ -106,7 +124,10 @@ class Llama:
             initialize_model_parallel(model_parallel_size)
 
         local_rank = int(os.environ.get("LOCAL_RANK", 0))
+        if device == "cuda":
             torch.cuda.set_device(local_rank)
+        elif device == "xpu":
+            torch.xpu.set_device(local_rank)
 
         # seed must be the same in all processes
         if config.torch_seed is not None:
@@ -189,10 +210,17 @@ class Llama:
                     "Currently int4 and fp8 are the only supported quantization methods."
                 )
         else:
+            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)
+            else:
+                torch.set_default_device(device)
+                if device == "xpu" and 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 +228,8 @@ 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)
 
@@ -266,14 +296,14 @@ 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
@@ -285,12 +315,10 @@ 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(
+                position_ids = torch.arange(prev_pos, cur_pos, dtype=torch.long)
-                    prev_pos, cur_pos, dtype=torch.long, device="cuda"
-                )
                 logits = self.model.forward(
                     position_ids,
                     tokens,