updated doc

docs/source/basic_tutorials/fp8_kv_cache.md

@@ -21,8 +21,8 @@ E4M3 offers higher precision for representing floating point numbers. However, d
 
 ## Current Hardware Support
 
-* Nvidia GPUs:  Supports both FP8E4M3 and FP8E5M2.
+* Nvidia GPUs:  Supports both FP8E4M3 (fp8) and FP8E5M2 (fp8_e5m2).
-* AMD GPUs: Supports FP8E4M3.
+* AMD GPUs: Supports FP8E4M3FNUZ (fp8).
 
 ## FP8 E5M2 KV Cache
 Example usage:

docs/source/basic_tutorials/launcher.md

@@ -91,6 +91,8 @@ Options:
 ## KV_CACHE_DTYPE
 ```shell
       --kv-cache-dtype <KV_CACHE_DTYPE>
+          Specify the data type for KV cache. By default, it uses the model's data type. CUDA 11.8+ supports `fp8(fp8_e4m3)` and 'fp8_e5m2', while ROCm (AMD GPU) supports `fp8(fp8_e4m3fnuz)'. If 'fp8' is chosen, a model checkpoint with scales for the KV cache should be provided. If not provided, the KV cache scaling factors default to 1.0, which may impact accuracy."
+          
           [env: KV_CACHE_DTYPE=]
           [possible values: fp8, fp8_e5m2]
 

server/text_generation_server/models/flash_causal_lm.py

@@ -729,6 +729,7 @@ class FlashCausalLM(Model):
         rank: int = 0,
         world_size: int = 1,
         sliding_window: Optional[int] = None,
+        kv_cache_dtype: Optional[torch.dtype] = None,
     ):
         self.num_layers = num_layers
         self.num_kv_heads = num_kv_heads
@@ -737,6 +738,8 @@ class FlashCausalLM(Model):
         self.cuda_graphs = {}
         self.kv_cache = []
 
+        self.kv_cache_dtype = kv_cache_dtype if kv_cache_dtype else dtype
+
         super(FlashCausalLM, self).__init__(
             model=model,
             tokenizer=tokenizer,
@@ -854,7 +857,7 @@ class FlashCausalLM(Model):
                 self.num_layers,
                 self.num_kv_heads,
                 self.head_size,
-                self.dtype,
+                self.kv_cache_dtype,
                 self.device,
             )
             max_bt = batch.max_blocks
@@ -873,7 +876,7 @@ class FlashCausalLM(Model):
 
         # Inspired by the original implementation in [vllm](https://github.com/vllm-project/vllm)
         # Calculate the number of blocks that can be allocated with the free memory
-        dtype_size = torch.tensor([], dtype=self.dtype).element_size()
+        dtype_size = torch.tensor([], dtype=self.kv_cache_dtype).element_size()
         cache_block_size = BLOCK_SIZE * self.num_kv_heads * self.head_size
         total_cache_size = self.num_layers * cache_block_size * 2 * dtype_size
 
@@ -893,7 +896,7 @@ class FlashCausalLM(Model):
             self.num_layers,
             self.num_kv_heads,
             self.head_size,
-            self.dtype,
+            self.kv_cache_dtype,
             self.device,
         )
 

server/text_generation_server/models/flash_llama.py

@@ -84,8 +84,9 @@ class FlashLlama(FlashCausalLM):
             num_layers=len(model.model.layers),
             num_kv_heads=model.model.num_key_value_heads,
             head_size=model.model.head_size,
-            dtype=torch.uint8 if "fp8" in kv_cache_dtype else dtype,
+            dtype=dtype,
             device=device,
             rank=rank,
             world_size=world_size,
+            kv_cache_dtype=torch.uint8 if "fp8" in kv_cache_dtype else dtype,
         )