feat: add cuda memory fraction (#659)

Close #673

integration-tests/models/test_flash_llama_gptq.py

@@ -48,8 +48,12 @@ async def test_flash_llama_gptq_all_params(flash_llama_gptq, response_snapshot):
 
 @pytest.mark.asyncio
 @pytest.mark.private
-async def test_flash_llama_gptq_load(flash_llama_gptq, generate_load, response_snapshot):
-    responses = await generate_load(flash_llama_gptq, "Test request", max_new_tokens=10, n=4)
+async def test_flash_llama_gptq_load(
+    flash_llama_gptq, generate_load, response_snapshot
+):
+    responses = await generate_load(
+        flash_llama_gptq, "Test request", max_new_tokens=10, n=4
+    )
 
     assert len(responses) == 4
     assert all([r.generated_text == responses[0].generated_text for r in responses])

integration-tests/models/test_flash_starcoder_gptq.py

@@ -17,7 +17,9 @@ async def flash_starcoder_gptq(flash_starcoder_gptq_handle):
 @pytest.mark.private
 async def test_flash_starcoder_gptq(flash_starcoder_gptq, response_snapshot):
     response = await flash_starcoder_gptq.generate(
-        "def geometric_mean(L: List[float]):", max_new_tokens=20, decoder_input_details=True,
+        "def geometric_mean(L: List[float]):",
+        max_new_tokens=20,
+        decoder_input_details=True,
     )
     assert response.details.generated_tokens == 20
     assert response == response_snapshot
@@ -25,7 +27,9 @@ async def test_flash_starcoder_gptq(flash_starcoder_gptq, response_snapshot):
 
 @pytest.mark.asyncio
 @pytest.mark.private
-async def test_flash_starcoder_gptq_default_params(flash_starcoder_gptq, response_snapshot):
+async def test_flash_starcoder_gptq_default_params(
+    flash_starcoder_gptq, response_snapshot
+):
     response = await flash_starcoder_gptq.generate(
         "def geometric_mean(L: List[float]):",
         max_new_tokens=20,
@@ -40,10 +44,17 @@ async def test_flash_starcoder_gptq_default_params(flash_starcoder_gptq, respons
 
 @pytest.mark.asyncio
 @pytest.mark.private
-async def test_flash_starcoder_gptq_load(flash_starcoder_gptq, generate_load, response_snapshot):
-    responses = await generate_load(flash_starcoder_gptq, "def geometric_mean(L: List[float]):", max_new_tokens=10, n=4)
+async def test_flash_starcoder_gptq_load(
+    flash_starcoder_gptq, generate_load, response_snapshot
+):
+    responses = await generate_load(
+        flash_starcoder_gptq,
+        "def geometric_mean(L: List[float]):",
+        max_new_tokens=10,
+        n=4,
+    )
 
     assert len(responses) == 4
     assert all([r.generated_text == responses[0].generated_text for r in responses])
 
-    assert responses == response_snapshot
+    assert responses == response_snapshot

launcher/src/main.rs

@@ -245,6 +245,11 @@ struct Args {
     #[clap(long, env)]
     disable_custom_kernels: bool,
 
+    /// Limit the CUDA available memory.
+    /// The allowed value equals the total visible memory multiplied by cuda-memory-fraction.
+    #[clap(default_value = "1.0", long, env)]
+    cuda_memory_fraction: f32,
+
     /// Outputs the logs in JSON format (useful for telemetry)
     #[clap(long, env)]
     json_output: bool,
@@ -299,6 +304,7 @@ fn shard_manager(
     disable_custom_kernels: bool,
     watermark_gamma: Option<f32>,
     watermark_delta: Option<f32>,
+    cuda_memory_fraction: f32,
     otlp_endpoint: Option<String>,
     status_sender: mpsc::Sender<ShardStatus>,
     shutdown: Arc<AtomicBool>,
@@ -368,6 +374,12 @@ fn shard_manager(
     envs.push(("MASTER_PORT".into(), master_port.to_string().into()));
     envs.push(("NCCL_ASYNC_ERROR_HANDLING".into(), "1".into()));
 
+    // CUDA memory fraction
+    envs.push((
+        "CUDA_MEMORY_FRACTION".into(),
+        cuda_memory_fraction.to_string().into(),
+    ));
+
     // Safetensors load fast
     envs.push(("SAFETENSORS_FAST_GPU".into(), "1".into()));
 
@@ -771,6 +783,7 @@ fn spawn_shards(
         let disable_custom_kernels = args.disable_custom_kernels;
         let watermark_gamma = args.watermark_gamma;
         let watermark_delta = args.watermark_delta;
+        let cuda_memory_fraction = args.cuda_memory_fraction;
         thread::spawn(move || {
             shard_manager(
                 model_id,
@@ -788,6 +801,7 @@ fn spawn_shards(
                 disable_custom_kernels,
                 watermark_gamma,
                 watermark_delta,
+                cuda_memory_fraction,
                 otlp_endpoint,
                 status_sender,
                 shutdown,

router/client/src/sharded_client.rs

@@ -101,8 +101,12 @@ impl ShardedClient {
             .iter_mut()
             .map(|client| Box::pin(client.warmup(max_input_length, max_prefill_tokens)))
             .collect();
-        // all shards return the same message
-        join_all(futures).await.pop().unwrap()
+        // Take the minimum value
+        let results = join_all(futures)
+            .await
+            .into_iter()
+            .collect::<Result<Vec<Option<u32>>>>()?;
+        Ok(results.into_iter().flatten().min())
     }
 
     /// Generate one token for each request in the given batch

server/exllama_kernels/setup.py

@@ -11,7 +11,7 @@ setup(
                 "exllama_kernels/cuda_buffers.cu",
                 "exllama_kernels/cuda_func/column_remap.cu",
                 "exllama_kernels/cuda_func/q4_matmul.cu",
-                "exllama_kernels/cuda_func/q4_matrix.cu"
+                "exllama_kernels/cuda_func/q4_matrix.cu",
             ],
         )
     ],

server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py

@@ -20,6 +20,7 @@ from text_generation_server.utils.layers import (
 )
 from safetensors import SafetensorError
 
+
 def load_multi_mqa(
     config, prefix: str, weights, bias: bool, head_size, num_heads, hidden_size
 ):
@@ -78,6 +79,7 @@ def _load_multi_mqa_gptq(
         bits, groupsize = weights._get_gptq_qparams()
 
         from text_generation_server.utils.layers import HAS_EXLLAMA
+
         use_exllama = HAS_EXLLAMA
         weight = (qweight, qzeros, scales, g_idx, bits, groupsize, use_exllama)
 

server/text_generation_server/models/flash_causal_lm.py

@@ -19,6 +19,7 @@ from text_generation_server.models.types import (
 )
 from text_generation_server.pb import generate_pb2
 from text_generation_server.utils import StoppingCriteria, HeterogeneousNextTokenChooser
+from text_generation_server.utils.dist import MEMORY_FRACTION
 
 tracer = trace.get_tracer(__name__)
 
@@ -738,7 +739,12 @@ class FlashCausalLM(Model):
         cache_block_size = BLOCK_SIZE * self.num_kv_heads * self.head_size
         total_cache_size = self.num_layers * cache_block_size * 2 * dtype_size
 
-        free_memory, _ = torch.cuda.mem_get_info(self.device)
+        total_free_memory, _ = torch.cuda.mem_get_info(self.device)
+        total_gpu_memory = torch.cuda.get_device_properties(self.device).total_memory
+
+        free_memory = max(
+            0, total_free_memory - (1 - MEMORY_FRACTION) * total_gpu_memory
+        )
 
         num_blocks = (
             int(free_memory // total_cache_size)

server/text_generation_server/models/model.py

@@ -10,6 +10,7 @@ from text_generation_server.pb.generate_pb2 import InfoResponse
 
 B = TypeVar("B", bound=Batch)
 
+
 class Model(ABC):
     def __init__(
         self,
@@ -21,9 +22,6 @@ class Model(ABC):
         rank: int = 0,
         world_size: int = 1,
     ):
-        if torch.cuda.is_available():
-            torch.cuda.set_per_process_memory_fraction(1.0)
-
         self.model = model.eval()
         self.tokenizer = tokenizer
         self.all_special_ids = set(tokenizer.all_special_ids)

server/text_generation_server/server.py

@@ -16,7 +16,6 @@ from text_generation_server.pb import generate_pb2_grpc, generate_pb2
 from text_generation_server.tracing import UDSOpenTelemetryAioServerInterceptor
 
 
-
 class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
     def __init__(self, model: Model, cache: Cache, server_urls: List[str]):
         self.cache = cache
@@ -146,7 +145,10 @@ def serve(
                 # When using GPTQ, Exllama kernels need some global kernels
                 # For which we have the finale shapes only after the model has loaded
                 # This will allocate those buffers.
-                from text_generation_server.utils.gptq.exllama import create_exllama_buffers
+                from text_generation_server.utils.gptq.exllama import (
+                    create_exllama_buffers,
+                )
+
                 create_exllama_buffers()
             except ImportError:
                 pass

server/text_generation_server/utils/dist.py

@@ -4,6 +4,13 @@ import torch
 from datetime import timedelta
 from loguru import logger
 
+# Tensor Parallelism settings
+RANK = int(os.getenv("RANK", "0"))
+WORLD_SIZE = int(os.getenv("WORLD_SIZE", "1"))
+
+# CUDA memory fraction
+MEMORY_FRACTION = float(os.getenv("CUDA_MEMORY_FRACTION", "1.0"))
+
 
 class FakeBarrier:
     def wait(self):
@@ -37,16 +44,14 @@ class FakeGroup:
 
 
 def initialize_torch_distributed():
-    rank = int(os.getenv("RANK", "0"))
-    world_size = int(os.getenv("WORLD_SIZE", "1"))
-
     if torch.cuda.is_available():
         from torch.distributed import ProcessGroupNCCL
 
         # Set the device id.
-        assert world_size <= torch.cuda.device_count(), "Each process is one gpu"
-        device = rank % torch.cuda.device_count()
+        assert WORLD_SIZE <= torch.cuda.device_count(), "Each process is one gpu"
+        device = RANK % torch.cuda.device_count()
         torch.cuda.set_device(device)
+        torch.cuda.set_per_process_memory_fraction(MEMORY_FRACTION, device)
         backend = "nccl"
         options = ProcessGroupNCCL.Options()
         options.is_high_priority_stream = True
@@ -55,22 +60,22 @@ def initialize_torch_distributed():
         backend = "gloo"
         options = None
 
-    if world_size == 1:
-        return FakeGroup(rank, world_size), rank, world_size
+    if WORLD_SIZE == 1:
+        return FakeGroup(RANK, WORLD_SIZE), RANK, WORLD_SIZE
     else:
         if os.getenv("DEBUG", None) == "1":
-            return FakeGroup(rank, world_size), rank, world_size
+            return FakeGroup(RANK, WORLD_SIZE), RANK, WORLD_SIZE
 
         if not torch.distributed.is_initialized():
             # Call the init process.
             torch.distributed.init_process_group(
                 backend=backend,
-                world_size=world_size,
-                rank=rank,
+                world_size=WORLD_SIZE,
+                rank=RANK,
                 timeout=timedelta(seconds=60),
                 pg_options=options,
             )
         else:
             logger.warning("torch.distributed is already initialized.")
 
-        return torch.distributed.group.WORLD, rank, world_size
+        return torch.distributed.group.WORLD, RANK, WORLD_SIZE

server/text_generation_server/utils/gptq/exllama.py

@@ -1,28 +1,28 @@
-
 import torch
 from exllama_kernels import make_q4, q4_matmul, prepare_buffers, set_tuning_params
 
 # Dummy tensor to pass instead of g_idx since there is no way to pass "None" to a C++ extension
-none_tensor = torch.empty((1, 1), device = "meta")
+none_tensor = torch.empty((1, 1), device="meta")
+
 
 def ext_make_q4(qweight, qzeros, scales, g_idx, device):
     """Construct Q4Matrix, return handle"""
-    return make_q4(qweight,
-                   qzeros,
-                   scales,
-                   g_idx if g_idx is not None else none_tensor,
-                   device)
+    return make_q4(
+        qweight, qzeros, scales, g_idx if g_idx is not None else none_tensor, device
+    )
+
 
 def ext_q4_matmul(x, q4, q4_width):
     """Matrix multiplication, returns x @ q4"""
     outshape = x.shape[:-1] + (q4_width,)
     x = x.view(-1, x.shape[-1])
-    output = torch.empty((x.shape[0], q4_width), dtype = torch.float16, device = x.device)
+    output = torch.empty((x.shape[0], q4_width), dtype=torch.float16, device=x.device)
 
     q4_matmul(x, q4, output)
 
     return output.view(outshape)
 
+
 MAX_DQ = 1
 MAX_INNER = 1
 ACT_ORDER = False
@@ -31,9 +31,10 @@ DEVICE = None
 TEMP_STATE = None
 TEMP_DQ = None
 
+
 def create_exllama_buffers():
     global MAX_DQ, MAX_INNER, ACT_ORDER, DEVICE, TEMP_STATE, TEMP_DQ
-                            
+
     if ACT_ORDER:
         # TODO: this should be set to rust side `max_total_tokens`, but TGI
         # does not offer an API to expose this variable to python, as this variable
@@ -45,7 +46,9 @@ def create_exllama_buffers():
         max_total_tokens = 1
 
     # This temp_state buffer is required to reorder X in the act-order case.
-    temp_state = torch.zeros((max_total_tokens, MAX_INNER), dtype=torch.float16, device=DEVICE)
+    temp_state = torch.zeros(
+        (max_total_tokens, MAX_INNER), dtype=torch.float16, device=DEVICE
+    )
     temp_dq = torch.zeros((1, MAX_DQ), dtype=torch.float16, device=DEVICE)
 
     # This temp_dq buffer is required to dequantize weights when using cuBLAS, typically for the prefill.
@@ -56,10 +59,12 @@ def create_exllama_buffers():
     matmul_no_half2 = False
     set_tuning_params(matmul_recons_thd, matmul_fused_remap, matmul_no_half2)
 
-    TEMP_STATE, TEMP_DQ =  temp_state, temp_dq
+    TEMP_STATE, TEMP_DQ = temp_state, temp_dq
+
 
 class Ex4bitLinear:
     """Linear layer implementation with per-group 4-bit quantization of the weights"""
+
     def __init__(self, qweight, qzeros, scales, g_idx, bias, bits, groupsize):
         global MAX_DQ, MAX_INNER, ACT_ORDER, DEVICE
         assert bits == 4
@@ -70,20 +75,24 @@ class Ex4bitLinear:
         self.scales = scales
         self.g_idx = g_idx.cpu() if g_idx is not None else None
         self.bias = bias if bias is not None else None
-        
-        if self.g_idx is not None and ((self.g_idx == 0).all() or torch.equal(g_idx.cpu(), torch.tensor([i // groupsize for i in range(g_idx.shape[0])], dtype=torch.int32))):
+
+        if self.g_idx is not None and (
+            (self.g_idx == 0).all()
+            or torch.equal(
+                g_idx.cpu(),
+                torch.tensor(
+                    [i // groupsize for i in range(g_idx.shape[0])], dtype=torch.int32
+                ),
+            )
+        ):
             self.empty_g_idx = True
             self.g_idx = None
-        
+
         assert self.device.type == "cuda"
         assert self.device.index is not None
 
         self.q4 = ext_make_q4(
-            self.qweight,
-            self.qzeros,
-            self.scales,
-            self.g_idx,
-            self.device.index
+            self.qweight, self.qzeros, self.scales, self.g_idx, self.device.index
         )
 
         self.height = qweight.shape[0] * 8
@@ -99,7 +108,8 @@ class Ex4bitLinear:
 
         # Handle act-order matrix
         if self.g_idx is not None:
-            if self.groupsize is None: raise ValueError("Found group index but no groupsize. What do?")
+            if self.groupsize is None:
+                raise ValueError("Found group index but no groupsize. What do?")
             self.act_order = True
         else:
             self.act_order = False
@@ -112,7 +122,7 @@ class Ex4bitLinear:
             MAX_INNER = max(MAX_INNER, self.height, self.width)
 
             ACT_ORDER = True
-    
+
     def forward(self, x):
         out = ext_q4_matmul(x, self.q4, self.width)
 

server/text_generation_server/utils/gptq/quantize.py

@@ -815,11 +815,7 @@ def load_weights_pre_hook(module_name, weights, recursive=False):
                 tensor = current_tensor.to(device=torch.device("cuda:0"))
                 if current_tensor.requires_grad:
                     tensor = nn.Parameter(tensor)
-                setdeepattr(
-                    module,
-                    local_param,
-                    tensor
-                )
+                setdeepattr(module, local_param, tensor)
 
     return inner
 

server/text_generation_server/utils/layers.py

@@ -17,9 +17,10 @@ except ImportError:
 from accelerate import init_empty_weights
 
 from text_generation_server.utils.gptq.quant_linear import QuantLinear
+
 HAS_EXLLAMA = True
 if os.getenv("DISABLE_EXLLAMA") == "True":
-    HAS_EXLLAMA=False
+    HAS_EXLLAMA = False
 try:
     from text_generation_server.utils.gptq.exllama import Ex4bitLinear
 except ImportError:

server/text_generation_server/utils/weights.py

@@ -146,7 +146,16 @@ class Weights:
             if self.process_group.size() > 1:
                 g_idx = self.get_tensor(f"{prefix}.g_idx")
                 if g_idx is not None:
-                    if not torch.equal(g_idx.cpu(), torch.tensor([i // groupsize for i in range(g_idx.shape[0])], dtype=torch.int32)) and not (g_idx == 0).all():
+                    if (
+                        not torch.equal(
+                            g_idx.cpu(),
+                            torch.tensor(
+                                [i // groupsize for i in range(g_idx.shape[0])],
+                                dtype=torch.int32,
+                            ),
+                        )
+                        and not (g_idx == 0).all()
+                    ):
                         # Exllama implementation does not support row tensor parallelism with act-order, as
                         # it would require to reorder input activations that are split unto several GPUs
                         use_exllama = False
@@ -154,18 +163,21 @@ class Weights:
             try:
                 qweight = self.get_sharded(f"{prefix}.qweight", dim=0)
             except RuntimeError:
-                raise RuntimeError("Cannot load `gptq` weight, make sure the model is already quantized, or quantize it with `text-generation-server quantize ORIGINAL_MODEL_ID NEW_MODEL_ID`")
-            
+                raise RuntimeError(
+                    "Cannot load `gptq` weight, make sure the model is already quantized, or quantize it with `text-generation-server quantize ORIGINAL_MODEL_ID NEW_MODEL_ID`"
+                )
 
             from text_generation_server.utils.layers import HAS_EXLLAMA
+
             if use_exllama:
                 if not HAS_EXLLAMA:
-                    logger.warning("Exllama GPTQ cuda kernels (which are faster) could have been used, but are not currently installed, try using BUILD_EXTENSIONS=True")
+                    logger.warning(
+                        "Exllama GPTQ cuda kernels (which are faster) could have been used, but are not currently installed, try using BUILD_EXTENSIONS=True"
+                    )
                     use_exllama = False
                 else:
                     logger.info("Using exllama kernels")
 
-
             if use_exllama:
                 if groupsize >= 0:
                     # Exllama reorders the weights in advance and the activations on the fly, thus
@@ -173,7 +185,9 @@ class Weights:
                     qzeros = self.get_sharded(f"{prefix}.qzeros", dim=0)
                     scales = self.get_sharded(f"{prefix}.scales", dim=0)
                 else:
-                    raise RuntimeError("Using exllama GPTQ kernel with groupsize<1 is not supported") 
+                    raise RuntimeError(
+                        "Using exllama GPTQ kernel with groupsize<1 is not supported"
+                    )
                     # qzeros = self.get_tensor(f"{prefix}.qzeros")
                     # scales = self.get_tensor(f"{prefix}.scales")