[XPU] Support W4A8C8-TP4-300B Model #4068
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Thanks for your contribution! |
zhupengyang
reviewed
Sep 11, 2025
|
|
iosmers
force-pushed
the
new_support_w4a8
branch
from
027483b to
aeb4eb7
Compare
iosmers
force-pushed
the
new_support_w4a8
branch
from
aeb4eb7 to
dd28881
Compare
iosmers
force-pushed
the
new_support_w4a8
branch
from
d5ce629 to
01908db
Compare
iosmers
force-pushed
the
new_support_w4a8
branch
from
b4fc62a to
3f64940
Compare
zhupengyang
reviewed
Sep 26, 2025
fastdeploy/model_executor/layers/backends/xpu/quantization/kv_cache.py
Outdated
Show resolved
Hide resolved
| PD_CHECK(ret == 0); | ||
| return {out, scale}; | ||
| } else { | ||
| } |
|
|
||
| if fd_config.quant_config and hasattr(fd_config.quant_config, "kv_cache_quant_type"): | ||
| self.quant_method: QuantMethodBase = fd_config.quant_config.get_quant_method(self) | ||
| print(f"quant_method: {self.quant_method}") |
| default_initializer=paddle.nn.initializer.Constant(0), | ||
| ) | ||
|
|
||
| def calculate_md5(self, arr): |
| XPU compute Fused MoE. | ||
| """ | ||
| from fastdeploy.model_executor.ops.xpu import xpu_moe_layer | ||
| # from fastdeploy.model_executor.ops.xpu import xpu_moe_layer |
| or self.quant_type == XPUKvCacheQuantzationTypes.FP8_ZP | ||
| or self.quant_type == XPUKvCacheQuantzationTypes.BLOCK_WISE_FP8 | ||
| ): | ||
| self.max_bound = 448.0 |
| ): | ||
| self.max_bound = 448.0 | ||
| elif self.quant_type == XPUKvCacheQuantzationTypes.INT4_ZP: | ||
| self.max_bound = 7.0 |
| cache_v_scale_tensor = get_tensor(state_dict.pop(self.cache_v_scale_name)).cast("float32").reshape_([-1]) | ||
|
|
||
| if self.cache_quant_config.has_zero_point: # cache_int4_zp | ||
| cache_k_scale = 1.0 / cache_k_scale_tensor |
| use for loader v1 | ||
| """ | ||
| if layer.cache_k_scale._is_initialized(): | ||
| layer.cache_k_out_scale.set_value(1 / layer.cache_k_scale) |
hong19860320
previously approved these changes
Sep 28, 2025
Collaborator
hong19860320
left a comment
hong19860320
left a comment
There was a problem hiding this comment.
LGTM, 先合入一版，有问题在后续PR补上吧
EmmonsCurse
approved these changes
Oct 10, 2025
DDDivano
approved these changes
Oct 10, 2025
qingqing01
approved these changes
Oct 10, 2025
hong19860320
approved these changes
Oct 10, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
9 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
本PR实现XPU适配ERNIE-4.5-300B-A47B-W4A8C8-TP4-Paddle模型
启动命令：
export XPU_VISIBLE_DEVICES="4,5,6,7"
python -m fastdeploy.entrypoints.openai.api_server --model ./PaddlePaddle/ERNIE-4.5-300B-A47B-W4A8C8-TP4-Paddle --port 8188
--tensor-parallel-size 4 --max-model-len 32768 --max-num-seqs 1 --quantization "W4A8" --gpu-memory-utilization 0.9