Support eetq weight only quantization (#1068)

# What does this PR do?

<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)


## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?


## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @


@OlivierDehaene OR @Narsil

 -->

---------

Co-authored-by: zhaosida <zhaosida@corp.netease.com>

Cargo.lock

@@ -2896,18 +2896,18 @@ dependencies = [
 
 [[package]]
 name = "thiserror"
-version = "1.0.48"
+version = "1.0.49"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d6d7a740b8a666a7e828dd00da9c0dc290dff53154ea77ac109281de90589b7"
+checksum = "1177e8c6d7ede7afde3585fd2513e611227efd6481bd78d2e82ba1ce16557ed4"
 dependencies = [
  "thiserror-impl",
 ]
 
 [[package]]
 name = "thiserror-impl"
-version = "1.0.48"
+version = "1.0.49"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "49922ecae66cc8a249b77e68d1d0623c1b2c514f0060c27cdc68bd62a1219d35"
+checksum = "10712f02019e9288794769fba95cd6847df9874d49d871d062172f9dd41bc4cc"
 dependencies = [
  "proc-macro2",
  "quote",

launcher/src/main.rs

@@ -21,11 +21,32 @@ mod env_runtime;
 
 #[derive(Clone, Copy, Debug, ValueEnum)]
 enum Quantization {
-    Bitsandbytes,
-    BitsandbytesNF4,
-    BitsandbytesFP4,
-    Gptq,
+    /// 4 bit quantization. Requires a specific GTPQ quantized model:
+    ///   https://hf.co/models?search=awq.
+    /// Should replace GPTQ models whereever possible because of the better latency
     Awq,
+    /// 8 bit quantization, doesn't require specific model.
+    /// Should be a drop-in replacement to bitsandbytes with much better performance.
+    /// Kernels are from https://github.com/NetEase-FuXi/EETQ.git
+    Eetq,
+    /// 4 bit quantization. Requires a specific GTPQ quantized model: https://hf.co/models?search=gptq.
+    /// text-generation-inference will use exllama (faster) kernels whereever possible, and use
+    /// triton kernel (wider support) when it's not.
+    /// AWQ has faster kernels.
+    Gptq,
+    /// Bitsandbytes 8bit. Can be applied on any model, will cut the memory requirement in half,
+    /// but it is known that the model will be much slower to run than the native f16.
+    #[deprecated(
+        since = "1.1.0",
+        note = "Use `eetq` instead, which provides better latencies overall and is drop-in in most cases"
+    )]
+    Bitsandbytes,
+    /// Bitsandbytes 4bit. Can be applied on any model, will cut the memory requirement by 4x,
+    /// but it is known that the model will be much slower to run than the native f16.
+    BitsandbytesNF4,
+    /// Bitsandbytes 4bit. nf4 should be preferred in most cases but maybe this one has better
+    /// perplexity performance for you model
+    BitsandbytesFP4,
 }
 
 impl std::fmt::Display for Quantization {
@@ -47,6 +68,9 @@ impl std::fmt::Display for Quantization {
             Quantization::Awq => {
                 write!(f, "awq")
             }
+            Quantization::Eetq => {
+                write!(f, "eetq")
+            }
         }
     }
 }
@@ -127,9 +151,7 @@ struct Args {
     #[clap(long, env)]
     num_shard: Option<usize>,
 
-    /// Whether you want the model to be quantized. This will use `bitsandbytes` for
-    /// quantization on the fly, or `gptq`. 4bit quantization is available through
-    /// `bitsandbytes` by providing the `bitsandbytes-fp4` or `bitsandbytes-nf4` options.
+    /// Whether you want the model to be quantized.
     #[clap(long, env, value_enum)]
     quantize: Option<Quantization>,
 

server/.gitignore

@@ -160,3 +160,4 @@ flash-attention/
 flash-attention-v2/
 vllm/
 llm-awq/
+eetq/

server/Makefile

@@ -2,6 +2,7 @@ include Makefile-flash-att
 include Makefile-flash-att-v2
 include Makefile-vllm
 include Makefile-awq
+include Makefile-eetq
 
 unit-tests:
 	pytest -s -vv -m "not private" tests

server/Makefile-eetq

@@ -0,0 +1,13 @@
+eetq_commit := 323827dd471458a84e9c840f614e4592b157a4b1
+
+eetq:
+    # Clone eetq
+	pip install packaging
+	git clone https://github.com/NetEase-FuXi/EETQ.git eetq
+
+build-eetq: eetq
+	cd eetq && git fetch && git checkout $(eetq_commit)
+	cd eetq && python setup.py build
+
+install-eetq: build-eetq
+	cd eetq && python setup.py install

server/text_generation_server/cli.py

@@ -18,6 +18,7 @@ class Quantization(str, Enum):
     bitsandbytes_fp4 = "bitsandbytes-fp4"
     gptq = "gptq"
     awq = "awq"
+    eetq = "eetq"
 
 
 class Dtype(str, Enum):

server/text_generation_server/utils/layers.py

@@ -5,6 +5,8 @@ import torch.distributed
 from torch import nn
 from torch.nn import functional as F
 from typing import List
+from loguru import logger
+from functools import lru_cache
 
 HAS_BITS_AND_BYTES = True
 try:
@@ -42,6 +44,13 @@ elif CAN_EXLLAMA:
 
 from typing import Optional
 
+HAS_EETQ = False
+try:
+    from EETQ import quant_weights, w8_a16_gemm
+    HAS_EETQ = True
+except ImportError:
+    pass
+
 # Monkey patching
 @classmethod
 def load_layer_norm(cls, prefix, weights, eps):
@@ -120,6 +129,30 @@ class FastLinear(nn.Module):
         return F.linear(input, self.weight, self.bias)
 
 
+class EETQLinear(nn.Module):
+    def __init__(
+        self,
+        weight,
+        bias,
+    ) -> None:
+        super().__init__()
+        device = weight.device
+        weight = torch.t(weight).contiguous().cpu()
+        weight, scale = quant_weights(weight, torch.int8, False)
+        if bias:
+            bias = weights.get_tensor(f"{prefix}.bias")
+        else:
+            bias = None
+        self.weight = weight.cuda(device)
+        self.scale = scale.cuda(device)
+        self.bias = bias.cuda(device) if bias is not None else None
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        output = w8_a16_gemm(input, self.weight, self.scale)
+        output = output + self.bias if self.bias is not None else output
+        return output
+
+
 class Linear8bitLt(nn.Module):
     def __init__(
         self,
@@ -211,10 +244,20 @@ class Linear4bit(nn.Module):
         return out
 
 
+@lru_cache(1)
+def warn_deprecate_bnb():
+    logger.warning("Bitsandbytes 8bit is deprecated, using `eetq` is a drop-in replacement, and has much better performnce")
+
 def get_linear(weight, bias, quantize):
     if quantize is None:
         linear = FastLinear(weight, bias)
+    elif quantize == "eetq":
+        if HAS_EETQ:
+            linear = EETQLinear(weight, bias)
+        else:
+            raise ImportError("Please install EETQ from https://github.com/NetEase-FuXi/EETQ")
     elif quantize == "bitsandbytes":
+        warn_deprecate_bnb()
         linear = Linear8bitLt(
             weight,
             bias,
@@ -298,8 +341,8 @@ class TensorParallelHead(SuperLayer):
             weight = weights.get_tensor(f"{prefix}.weight")
             should_gather = False
 
-        # GPTQ and AWQ don't quantize heads (nor embeddings)
-        if config.quantize in ["gptq", "awq"]:
+        # GPTQ,AWQ,EETQ don't quantize heads (nor embeddings)
+        if config.quantize in ["gptq", "awq", "eetq"]:
             quantize = None
         else:
             quantize = config.quantize