[Kernel] FP8 support for MoE kernel / Mixtral (#4244)

This PR is the first step towards fixing https://github.com/vllm-project/vllm/pull/3208

It implements dynamic per-tensor scaling (see https://github.com/vllm-project/vllm/pull/4118), so users do not need to compute activation scales on a calibration dataset and they also don't need to convert their model checkpoints. It is enough to specify the `quantization="fp8"` argument. You can try out the PR like this:

```python
from vllm import LLM, SamplingParams

prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

llm = LLM(model="mistralai/Mixtral-8x7B-Instruct-v0.1", tensor_parallel_size=2, quantization="fp8")

outputs = llm.generate(prompts, sampling_params)

# Print the outputs.
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```

**Performance**: For this PR, the focus is on making the code clean (while still trying to get reasonable performance), there is a bunch of optimizations that we will submit as a follow up PR that significantly improve the performance (similar to the numbers in https://github.com/vllm-project/vllm/pull/3954). With this PR, the results are as follows:

<img width="725" alt="Screenshot 2024-04-21 at 1 31 50 PM" src="https://github.com/vllm-project/vllm/assets/113316/d8fe1118-07a0-4d4e-8530-37a77d465a03">


**Accuracy**: The accuracy with this PR on MMLU on `mistralai/Mixtral-8x7B-v0.1` is as follows:

```
|      Groups      |Version|Filter|n-shot|Metric|Value |   |Stderr|
|------------------|-------|------|-----:|------|-----:|---|-----:|
|mmlu              |N/A    |none  |     0|acc   |0.7018|±  |0.0036|
| - humanities     |N/A    |none  |     5|acc   |0.6472|±  |0.0065|
| - other          |N/A    |none  |     5|acc   |0.7673|±  |0.0072|
| - social_sciences|N/A    |none  |     5|acc   |0.8099|±  |0.0070|
| - stem           |N/A    |none  |     5|acc   |0.6131|±  |0.0083|
```
this compares favorably with the fp16 results which are
```
|      Groups      |Version|Filter|n-shot|Metric|Value |   |Stderr|
|------------------|-------|------|-----:|------|-----:|---|-----:|
|mmlu              |N/A    |none  |     0|acc   |0.7020|±  |0.1313|
| - humanities     |N/A    |none  |     5|acc   |0.6425|±  |0.1349|
| - other          |N/A    |none  |     5|acc   |0.7744|±  |0.1038|
| - social_sciences|N/A    |none  |     5|acc   |0.8131|±  |0.0695|
| - stem           |N/A    |none  |     5|acc   |0.6108|±  |0.1383|
```

Happy hacking!
This commit is contained in:
Philipp Moritz
2024-04-24 01:18:23 +00:00
committed by GitHub
parent 1e8f4252aa
commit eace8bf0b9
10 changed files with 385 additions and 21 deletions
@@ -0,0 +1,146 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"2": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 4
},
"4": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"8": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"16": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"24": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"32": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 4
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 4
}
}
@@ -21,6 +21,8 @@ def fused_moe_kernel(
a_ptr,
b_ptr,
c_ptr,
a_scale_ptr,
b_scale_ptr,
topk_weights_ptr,
sorted_token_ids_ptr,
expert_ids_ptr,
@@ -49,6 +51,7 @@ def fused_moe_kernel(
MUL_ROUTED_WEIGHT: tl.constexpr,
top_k: tl.constexpr,
compute_type: tl.constexpr,
use_fp8: tl.constexpr,
):
"""
Implements the fused computation for a Mixture of Experts (MOE) using
@@ -111,6 +114,10 @@ def fused_moe_kernel(
b_ptrs = b_ptr + off_experts * stride_be + (offs_k[:, None] * stride_bk +
offs_bn[None, :] * stride_bn)
if use_fp8:
a_scale = tl.load(a_scale_ptr)
b_scale = tl.load(b_scale_ptr + off_experts)
# -----------------------------------------------------------
# Iterate to compute a block of the C matrix.
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
@@ -129,7 +136,10 @@ def fused_moe_kernel(
mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
other=0.0)
# We accumulate along the K dimension.
accumulator += tl.dot(a, b)
if use_fp8:
accumulator = tl.dot(a, b, acc=accumulator)
else:
accumulator += tl.dot(a, b)
# Advance the ptrs to the next K block.
a_ptrs += BLOCK_SIZE_K * stride_ak
b_ptrs += BLOCK_SIZE_K * stride_bk
@@ -140,7 +150,10 @@ def fused_moe_kernel(
other=0)
accumulator = accumulator * moe_weight[:, None]
accumulator = accumulator.to(compute_type)
if use_fp8:
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
else:
accumulator = accumulator.to(compute_type)
# -----------------------------------------------------------
# Write back the block of the output
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
@@ -207,15 +220,24 @@ def moe_align_block_size(
def invoke_fused_moe_kernel(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
topk_weights: torch.Tensor, topk_ids: torch.Tensor,
B_scale: torch.Tensor, topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
sorted_token_ids: torch.Tensor,
expert_ids: torch.Tensor,
num_tokens_post_padded: torch.Tensor,
mul_routed_weight: bool, top_k: int,
config: Dict[str, Any]) -> None:
config: Dict[str, Any], compute_type: tl.dtype,
use_fp8: bool) -> None:
assert topk_weights.stride(1) == 1
assert sorted_token_ids.stride(0) == 1
if not use_fp8:
A_scale = None
assert B_scale is None
else:
A, A_scale = ops.scaled_fp8_quant(A)
assert B_scale is not None
grid = lambda META: (triton.cdiv(sorted_token_ids.shape[0], META[
'BLOCK_SIZE_M']) * triton.cdiv(B.shape[1], META['BLOCK_SIZE_N']), )
@@ -223,6 +245,8 @@ def invoke_fused_moe_kernel(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
A,
B,
C,
A_scale,
B_scale,
topk_weights,
sorted_token_ids,
expert_ids,
@@ -240,18 +264,21 @@ def invoke_fused_moe_kernel(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
C.stride(2),
MUL_ROUTED_WEIGHT=mul_routed_weight,
top_k=top_k,
compute_type=tl.bfloat16 if A.dtype == torch.bfloat16 else tl.float16,
compute_type=compute_type,
use_fp8=use_fp8,
**config,
)
def get_config_file_name(E: int, N: int) -> str:
def get_config_file_name(E: int, N: int, dtype: Optional[str]) -> str:
device_name = torch.cuda.get_device_name().replace(" ", "_")
return f"E={E},N={N},device_name={device_name}.json"
dtype_selector = "" if not dtype else f",dtype={dtype}"
return f"E={E},N={N},device_name={device_name}{dtype_selector}.json"
@functools.lru_cache
def get_moe_configs(E: int, N: int) -> Optional[Dict[int, Any]]:
def get_moe_configs(E: int, N: int,
dtype: Optional[str]) -> Optional[Dict[int, Any]]:
"""
Return optimized configurations for the fused MoE kernel.
@@ -263,7 +290,7 @@ def get_moe_configs(E: int, N: int) -> Optional[Dict[int, Any]]:
# First look up if an optimized configuration is available in the configs
# directory
json_file_name = get_config_file_name(E, N)
json_file_name = get_config_file_name(E, N, dtype)
config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
@@ -288,6 +315,9 @@ def fused_moe(
renormalize: bool,
inplace: bool = False,
override_config: Optional[Dict[str, Any]] = None,
use_fp8: bool = False,
w1_scale: Optional[torch.Tensor] = None,
w2_scale: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
This function computes a Mixture of Experts (MoE) layer using two sets of
@@ -305,6 +335,12 @@ def fused_moe(
Defaults to False.
- override_config (Optional[Dict[str, Any]]): Optional override
for the kernel configuration.
- use_fp8 (bool): If True, use fp8 arithmetic to compute the inner
products for w1 and w2. Defaults to False.
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
w1.
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
w2.
Returns:
- torch.Tensor: The output tensor after applying the MoE layer.
@@ -358,7 +394,8 @@ def fused_moe(
config = override_config
else:
# First try to load optimal config from the file
configs = get_moe_configs(E, w2.shape[2])
configs = get_moe_configs(E, w2.shape[2],
"float8" if use_fp8 else None)
if configs:
# If an optimal configuration map has been found, look up the
@@ -394,17 +431,37 @@ def fused_moe(
sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size(
topk_ids, config['BLOCK_SIZE_M'], E)
invoke_fused_moe_kernel(hidden_states, w1, intermediate_cache1,
topk_weights, topk_ids, sorted_token_ids,
expert_ids, num_tokens_post_padded, False,
topk_ids.shape[1], config)
invoke_fused_moe_kernel(hidden_states,
w1,
intermediate_cache1,
w1_scale,
topk_weights,
topk_ids,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
False,
topk_ids.shape[1],
config,
compute_type=tl.float16,
use_fp8=use_fp8)
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
invoke_fused_moe_kernel(intermediate_cache2, w2, intermediate_cache3,
topk_weights, topk_ids, sorted_token_ids,
expert_ids, num_tokens_post_padded, True, 1,
config)
invoke_fused_moe_kernel(intermediate_cache2,
w2,
intermediate_cache3,
w2_scale,
topk_weights,
topk_ids,
sorted_token_ids,
expert_ids,
num_tokens_post_padded,
True,
1,
config,
compute_type=tl.float16,
use_fp8=use_fp8)
if inplace:
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape),