Skip to content

feat(models): add qwen3.5-35b-a3b-fp8 model config and HuggingFace mapping #5076

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
snehalv2002 wants to merge 1 commit into
base: pr/fp8-orbax-restoration
Choose a base branch
from
Draft

feat(models): add qwen3.5-35b-a3b-fp8 model config and HuggingFace mapping #5076

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions src/maxtext/checkpoint_conversion/utils/hf_model_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1931,6 +1931,8 @@ def __init__(self, **kwargs):
"qwen3-next-80b-a3b": qwen3_next_80b_a3b_config,
"qwen3.5-397b-a17b": qwen3_5_397b_a17b_config,
"qwen3.5-35b-a3b": qwen3_5_35b_a3b_config,
"qwen3.5-35b-a3b-fp8": qwen3_5_35b_a3b_config,
"qwen3.5-35b-fp8": qwen3_5_35b_a3b_config,
"mixtral-8x7b": mixtral_8x7b_config,
"mixtral-8x22b": mixtral_8x22b_config,
"olmo3-7b": olmo3_7b_config,
Expand Down
2 changes: 2 additions & 0 deletions src/maxtext/checkpoint_conversion/utils/hf_shape.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1317,6 +1317,8 @@ def DEEPSEEKV4_HF_WEIGHTS_TO_SHAPE(config):
"mixtral-8x7b": MIXTRAL_HF_WEIGHTS_TO_SHAPE,
"mixtral-8x22b": MIXTRAL_HF_WEIGHTS_TO_SHAPE,
"qwen3.5-35b-a3b": QWEN3_5_HF_WEIGHTS_TO_SHAPE,
"qwen3.5-35b-a3b-fp8": QWEN3_5_HF_WEIGHTS_TO_SHAPE,
"qwen3.5-35b-fp8": QWEN3_5_HF_WEIGHTS_TO_SHAPE,
Comment on lines +1320 to +1321

@gemini-code-assist gemini-code-assist Bot Aug 31, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The QWEN3_5_HF_WEIGHTS_TO_SHAPE function in hf_shape.py needs to be updated to define the shapes of the newly mapped weight_scale parameters (such as self_attn.q_proj.weight_scale, linear_attn.in_proj_qkv.weight_scale, etc.). Without these shape definitions, any attempt to convert checkpoints for the FP8 models (e.g., MaxText -> HF) will fail with a KeyError when looking up the expected shapes of the scale tensors.

"qwen3.5-397b-a17b": QWEN3_5_HF_WEIGHTS_TO_SHAPE,
"qwen3-next-80b-a3b": QWEN3_NEXT_HF_WEIGHTS_TO_SHAPE,
}
92 changes: 92 additions & 0 deletions src/maxtext/checkpoint_conversion/utils/param_mapping.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -912,6 +912,18 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
f"{prefix}-attention-attention-out-kernel": [
f"model.language_model.layers.{i}.self_attn.o_proj.weight" for i in hf_indices
],
f"{prefix}-attention-attention-query-kernel_scale": [
f"model.language_model.layers.{i}.self_attn.q_proj.weight_scale" for i in hf_indices
],
f"{prefix}-attention-attention-key-kernel_scale": [
f"model.language_model.layers.{i}.self_attn.k_proj.weight_scale" for i in hf_indices
],
f"{prefix}-attention-attention-value-kernel_scale": [
f"model.language_model.layers.{i}.self_attn.v_proj.weight_scale" for i in hf_indices
],
f"{prefix}-attention-attention-out-kernel_scale": [
f"model.language_model.layers.{i}.self_attn.o_proj.weight_scale" for i in hf_indices
],
f"{prefix}-attention-attention-query_norm-scale": [
f"model.language_model.layers.{i}.self_attn.q_norm.weight" for i in hf_indices
],
Expand Down Expand Up @@ -940,6 +952,20 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
)
for i in hf_indices
],
f"{prefix}-attention-in_proj_qkvz-kernel_scale": [
(
f"model.language_model.layers.{i}.linear_attn.in_proj_qkv.weight_scale",
f"model.language_model.layers.{i}.linear_attn.in_proj_z.weight_scale",
)
for i in hf_indices
],
f"{prefix}-attention-in_proj_ba-kernel_scale": [
(
f"model.language_model.layers.{i}.linear_attn.in_proj_b.weight_scale",
f"model.language_model.layers.{i}.linear_attn.in_proj_a.weight_scale",
)
for i in hf_indices
],
f"{prefix}-attention-conv1d-kernel": [
f"model.language_model.layers.{i}.linear_attn.conv1d.weight" for i in hf_indices
],
Expand All @@ -953,6 +979,9 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
f"{prefix}-attention-out_proj-kernel": [
f"model.language_model.layers.{i}.linear_attn.out_proj.weight" for i in hf_indices
],
f"{prefix}-attention-out_proj-kernel_scale": [
f"model.language_model.layers.{i}.linear_attn.out_proj.weight_scale" for i in hf_indices
],
}
)

Expand All @@ -974,6 +1003,18 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
f"{prefix}-mlp-shared_expert_gate-kernel": [
f"model.language_model.layers.{i}.mlp.shared_expert_gate.weight" for i in hf_indices
],
f"{prefix}-mlp-shared_expert-wi_0-kernel_scale": [
f"model.language_model.layers.{i}.mlp.shared_expert.gate_proj.weight_scale" for i in hf_indices
],
f"{prefix}-mlp-shared_expert-wi_1-kernel_scale": [
f"model.language_model.layers.{i}.mlp.shared_expert.up_proj.weight_scale" for i in hf_indices
],
f"{prefix}-mlp-shared_expert-wo-kernel_scale": [
f"model.language_model.layers.{i}.mlp.shared_expert.down_proj.weight_scale" for i in hf_indices
],
f"{prefix}-mlp-shared_expert_gate-kernel_scale": [
f"model.language_model.layers.{i}.mlp.shared_expert_gate.weight_scale" for i in hf_indices
],
}
)

Expand All @@ -986,6 +1027,12 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
(f"{prefix}-mlp-routed_experts-wi_0", f"{prefix}-mlp-routed_experts-wi_1"): [
f"model.language_model.layers.{i}.mlp.experts.gate_up_proj" for i in hf_indices
],
f"{prefix}-mlp-routed_experts-wo-kernel_scale": [
f"model.language_model.layers.{i}.mlp.experts.down_proj.weight_scale" for i in hf_indices
],
(f"{prefix}-mlp-routed_experts-wi_0-kernel_scale", f"{prefix}-mlp-routed_experts-wi_1-kernel_scale"): [
f"model.language_model.layers.{i}.mlp.experts.gate_up_proj.weight_scale" for i in hf_indices
],
}
)
else:
Expand All @@ -1009,6 +1056,10 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
f"{prefix}-attention-attention-key-kernel": f"model.language_model.layers.{i}.self_attn.k_proj.weight",
f"{prefix}-attention-attention-value-kernel": f"model.language_model.layers.{i}.self_attn.v_proj.weight",
f"{prefix}-attention-attention-out-kernel": f"model.language_model.layers.{i}.self_attn.o_proj.weight",
f"{prefix}-attention-attention-query-kernel_scale": f"model.language_model.layers.{i}.self_attn.q_proj.weight_scale",
f"{prefix}-attention-attention-key-kernel_scale": f"model.language_model.layers.{i}.self_attn.k_proj.weight_scale",
f"{prefix}-attention-attention-value-kernel_scale": f"model.language_model.layers.{i}.self_attn.v_proj.weight_scale",
f"{prefix}-attention-attention-out-kernel_scale": f"model.language_model.layers.{i}.self_attn.o_proj.weight_scale",
f"{prefix}-attention-attention-query_norm-scale": f"model.language_model.layers.{i}.self_attn.q_norm.weight",
f"{prefix}-attention-attention-key_norm-scale": f"model.language_model.layers.{i}.self_attn.k_norm.weight",
}
Expand All @@ -1027,11 +1078,20 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
f"model.language_model.layers.{i}.linear_attn.in_proj_b.weight",
f"model.language_model.layers.{i}.linear_attn.in_proj_a.weight",
),
f"{prefix}-attention-in_proj_qkvz-kernel_scale": (
f"model.language_model.layers.{i}.linear_attn.in_proj_qkv.weight_scale",
f"model.language_model.layers.{i}.linear_attn.in_proj_z.weight_scale",
),
f"{prefix}-attention-in_proj_ba-kernel_scale": (
f"model.language_model.layers.{i}.linear_attn.in_proj_b.weight_scale",
f"model.language_model.layers.{i}.linear_attn.in_proj_a.weight_scale",
),
f"{prefix}-attention-conv1d-kernel": f"model.language_model.layers.{i}.linear_attn.conv1d.weight",
f"{prefix}-attention-A_log": f"model.language_model.layers.{i}.linear_attn.A_log",
f"{prefix}-attention-dt_bias": f"model.language_model.layers.{i}.linear_attn.dt_bias",
f"{prefix}-attention-norm-rms_norm-scale": f"model.language_model.layers.{i}.linear_attn.norm.weight",
f"{prefix}-attention-out_proj-kernel": f"model.language_model.layers.{i}.linear_attn.out_proj.weight",
f"{prefix}-attention-out_proj-kernel_scale": f"model.language_model.layers.{i}.linear_attn.out_proj.weight_scale",
}
)

Expand All @@ -1045,6 +1105,10 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
f"{prefix}-mlp-shared_expert-wi_1-kernel": (f"{hf_mlp}.shared_expert.up_proj.weight"),
f"{prefix}-mlp-shared_expert-wo-kernel": (f"{hf_mlp}.shared_expert.down_proj.weight"),
f"{prefix}-mlp-shared_expert_gate-kernel": (f"{hf_mlp}.shared_expert_gate.weight"),
f"{prefix}-mlp-shared_expert-wi_0-kernel_scale": (f"{hf_mlp}.shared_expert.gate_proj.weight_scale"),
f"{prefix}-mlp-shared_expert-wi_1-kernel_scale": (f"{hf_mlp}.shared_expert.up_proj.weight_scale"),
f"{prefix}-mlp-shared_expert-wo-kernel_scale": (f"{hf_mlp}.shared_expert.down_proj.weight_scale"),
f"{prefix}-mlp-shared_expert_gate-kernel_scale": (f"{hf_mlp}.shared_expert_gate.weight_scale"),
}
)

Expand All @@ -1056,6 +1120,11 @@ def QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING(config, maxtext_config, scan_layers=Fals
f"{prefix}-mlp-routed_experts-wi_0",
f"{prefix}-mlp-routed_experts-wi_1",
): f"model.language_model.layers.{i}.mlp.experts.gate_up_proj",
f"{prefix}-mlp-routed_experts-wo-kernel_scale": f"model.language_model.layers.{i}.mlp.experts.down_proj.weight_scale",
(
f"{prefix}-mlp-routed_experts-wi_0-kernel_scale",
f"{prefix}-mlp-routed_experts-wi_1-kernel_scale",
): f"model.language_model.layers.{i}.mlp.experts.gate_up_proj.weight_scale",
}
)

Expand Down Expand Up @@ -1251,6 +1320,13 @@ def concat_ba_and_transpose(input_tensor, target_shape=None):
interleaved = np.concatenate([b_r, a_r], axis=1)
return interleaved.reshape(-1, b_m.shape[-1]).T

def reshape_scale(input_tensor, target_shape=None):
if target_shape is None:
return input_tensor
if input_tensor.ndim == 2:
return input_tensor.transpose().reshape(target_shape)
return input_tensor.reshape(target_shape)
Comment on lines +1323 to +1328

@gemini-code-assist gemini-code-assist Bot Aug 31, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

To support the composite scale keys (such as in_proj_qkvz-kernel_scale, in_proj_ba-kernel_scale, and the routed experts scale tuple), we need custom helper functions to handle splitting and concatenating these scale tensors. Using reshape_scale directly on composite keys will cause AttributeError or incorrect outputs because composite keys expect tuples of tensors when saving to HF, and receive tuples of tensors when loading from HF.

  def reshape_scale(input_tensor, target_shape=None):
    if target_shape is None:
      return input_tensor
    if input_tensor.ndim == 2:
      return input_tensor.transpose().reshape(target_shape)
    return input_tensor.reshape(target_shape)

  def process_wi_0_wi_1_scale(input_tensor, target_shape=None):
    if saving_to_hf:
      wi_0, wi_1 = input_tensor
      return np.concatenate([wi_0, wi_1], axis=-1)
    else:
      return np.split(input_tensor, 2, axis=-1)

  def split_qkvz_scale(input_tensor, target_shape=None):
    if saving_to_hf:
      conv_dim = 2 * H_k * D_k + H_v * D_v
      return input_tensor[:conv_dim], input_tensor[conv_dim:]
    else:
      qkv_scale, z_scale = input_tensor
      return np.concatenate([qkv_scale, z_scale], axis=0)

  def split_ba_scale(input_tensor, target_shape=None):
    if saving_to_hf:
      return input_tensor[:H_v], input_tensor[H_v:]
    else:
      b_scale, a_scale = input_tensor
      return np.concatenate([b_scale, a_scale], axis=0)


# Initialize Hooks
hooks = {
"params-decoder-logits_dense-kernel": transpose,
Expand All @@ -1272,23 +1348,35 @@ def concat_ba_and_transpose(input_tensor, target_shape=None):
if is_full_attention_layer:
for key in ["query", "key", "value", "out"]:
hooks[f"{prefix}-attention-attention-{key}-kernel"] = reshape_kernel # pyrefly: ignore[bad-assignment]
hooks[f"{prefix}-attention-attention-{key}-kernel_scale"] = reshape_scale
else:
hooks[f"{prefix}-attention-in_proj_qkvz-kernel"] = concat_qkvz_and_transpose
hooks[f"{prefix}-attention-in_proj_ba-kernel"] = concat_ba_and_transpose
hooks[f"{prefix}-attention-out_proj-kernel"] = transpose
hooks[f"{prefix}-attention-conv1d-kernel"] = permute_conv
hooks[f"{prefix}-attention-in_proj_qkvz-kernel_scale"] = reshape_scale
hooks[f"{prefix}-attention-in_proj_ba-kernel_scale"] = reshape_scale
Comment on lines +1357 to +1358

@gemini-code-assist gemini-code-assist Bot Aug 31, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

Use the newly defined split_qkvz_scale and split_ba_scale helper functions to correctly handle the composite scale keys.

Suggested change
hooks[f"{prefix}-attention-in_proj_qkvz-kernel_scale"] = reshape_scale
hooks[f"{prefix}-attention-in_proj_ba-kernel_scale"] = reshape_scale
hooks[f"{prefix}-attention-in_proj_qkvz-kernel_scale"] = split_qkvz_scale
hooks[f"{prefix}-attention-in_proj_ba-kernel_scale"] = split_ba_scale

hooks[f"{prefix}-attention-out_proj-kernel_scale"] = reshape_scale

mlp_prefix = f"{prefix}-mlp"
hooks[f"{mlp_prefix}-routed_experts-gate-kernel"] = transpose
hooks[f"{mlp_prefix}-shared_expert-wi_0-kernel"] = transpose
hooks[f"{mlp_prefix}-shared_expert-wi_1-kernel"] = transpose
hooks[f"{mlp_prefix}-shared_expert-wo-kernel"] = transpose
hooks[f"{mlp_prefix}-shared_expert_gate-kernel"] = transpose
hooks[f"{mlp_prefix}-shared_expert-wi_0-kernel_scale"] = reshape_scale
hooks[f"{mlp_prefix}-shared_expert-wi_1-kernel_scale"] = reshape_scale
hooks[f"{mlp_prefix}-shared_expert-wo-kernel_scale"] = reshape_scale
hooks[f"{mlp_prefix}-shared_expert_gate-kernel_scale"] = reshape_scale
# pyrefly: ignore[unsupported-operation]
hooks[(f"{mlp_prefix}-routed_experts-wi_0", f"{mlp_prefix}-routed_experts-wi_1")] = (
process_wi_0_wi_1 # pyrefly: ignore[unsupported-operation]
)
hooks[f"{mlp_prefix}-routed_experts-wo"] = transpose_expert
hooks[f"{mlp_prefix}-routed_experts-wo-kernel_scale"] = reshape_scale
hooks[(f"{mlp_prefix}-routed_experts-wi_0-kernel_scale", f"{mlp_prefix}-routed_experts-wi_1-kernel_scale")] = (
reshape_scale
)
Comment on lines +1377 to +1379

@gemini-code-assist gemini-code-assist Bot Aug 31, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

Use the newly defined process_wi_0_wi_1_scale helper function to correctly handle the composite routed experts scale key.

Suggested change
hooks[(f"{mlp_prefix}-routed_experts-wi_0-kernel_scale", f"{mlp_prefix}-routed_experts-wi_1-kernel_scale")] = (
reshape_scale
)
hooks[(f"{mlp_prefix}-routed_experts-wi_0-kernel_scale", f"{mlp_prefix}-routed_experts-wi_1-kernel_scale")] = (
process_wi_0_wi_1_scale
)


# Vision hooks for Qwen3.5
vision_config = config.get("vision_config", None)
Expand Down Expand Up @@ -4263,6 +4351,8 @@ def mhc_concat_scale(input_tensors, target_shape=None):
"qwen3-next-80b-a3b": QWEN3_NEXT_MAXTEXT_TO_HF_PARAM_MAPPING,
"qwen3.5-397b-a17b": QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING,
"qwen3.5-35b-a3b": QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING,
"qwen3.5-35b-a3b-fp8": QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING,
"qwen3.5-35b-fp8": QWEN3_5_MAXTEXT_TO_HF_PARAM_MAPPING,
"mixtral-8x7b": MIXTRAL_MAXTEXT_TO_HF_PARAM_MAPPING,
"mixtral-8x22b": MIXTRAL_MAXTEXT_TO_HF_PARAM_MAPPING,
"olmo3-7b": OLMO3_MAXTEXT_TO_HF_PARAM_MAPPING,
Expand Down Expand Up @@ -4317,6 +4407,8 @@ def mhc_concat_scale(input_tensors, target_shape=None):
"qwen3-omni-30b-a3b": QWEN3_OMNI_MOE_MAXTEXT_TO_HF_PARAM_HOOK_FN,
"qwen3.5-397b-a17b": QWEN3_5_MAXTEXT_TO_HF_PARAM_HOOK_FN,
"qwen3.5-35b-a3b": QWEN3_5_MAXTEXT_TO_HF_PARAM_HOOK_FN,
"qwen3.5-35b-a3b-fp8": QWEN3_5_MAXTEXT_TO_HF_PARAM_HOOK_FN,
"qwen3.5-35b-fp8": QWEN3_5_MAXTEXT_TO_HF_PARAM_HOOK_FN,
"qwen3-next-80b-a3b": QWEN3_NEXT_MAXTEXT_TO_HF_PARAM_HOOK_FN,
"mixtral-8x7b": MIXTRAL_MAXTEXT_TO_HF_PARAM_HOOK_FN,
"mixtral-8x22b": MIXTRAL_MAXTEXT_TO_HF_PARAM_HOOK_FN,
Expand Down
75 changes: 75 additions & 0 deletions src/maxtext/configs/models/qwen3.5-35b-a3b-fp8.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,75 @@
# Copyright 2026 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# model config for qwen3.5-35b-a3b-fp8 (FP8 weight-only storage with dynamic dequantization)

decoder_block: "qwen3_5"

# Core Architectural Parameters
base_emb_dim: 2048
base_num_decoder_layers: 40
base_num_query_heads: 16
base_num_kv_heads: 2
head_dim: 256
vocab_size: 248320
normalization_layer_epsilon: 1.0e-6

# MoE Specific Parameters
# Set base_mlp_dim to match base_moe_mlp_dim to pass validation for fully MoE models.
base_mlp_dim: 512
base_moe_mlp_dim: 512
num_experts: 256
shared_experts: 1
num_experts_per_tok: 8
norm_topk_prob: True

# GatedDeltaNet Specific Parameters for Linear Attention (GDN)
inhomogeneous_layer_cycle_interval: 4
gdn_conv_kernel_dim: 4
gdn_key_head_dim: 128
gdn_value_head_dim: 128
gdn_num_key_heads: 16
gdn_num_value_heads: 32
gdn_chunk_size: 64

# RoPE Settings
rope_max_timescale: 10000000
partial_rotary_factor: 0.25

# General Model Settings
enable_dropout: False

# Vision Encoder Configuration (need to set use_multimodal=true)
vision_encoder_block: "qwen3_5"
# Based on Qwen3.5 MoE Vision Model Config
image_size_for_vit: 768
hidden_size_for_vit: 1152
intermediate_size_for_vit: 4304
num_attention_heads_for_vit: 16
num_hidden_layers_for_vit: 27
num_channels_for_vit: 3
patch_size_for_vit: 16
temporal_patch_size_for_vit: 2
spatial_merge_size_for_vit: 2
out_hidden_size_for_vit: 2048 # Projects to decoder emb_dim (2048)
num_position_embeddings_for_vit: 2304
deepstack_visual_indexes_for_vit: [] # No deepstack for Qwen3.5 VL
rope_theta_for_vit: 10000

# MRoPE Settings (Multi-dimensional RoPE for multimodal)
use_mrope: true
mrope_section: [11, 11, 10]

weight_dtype: "float8_e4m3fn"
dtype: "bfloat16"
2 changes: 2 additions & 0 deletions src/maxtext/utils/globals.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,6 +85,8 @@
"qwen3-next-80b-a3b": "Qwen/Qwen3-Next-80B-A3B-Instruct",
"qwen3.5-397b-a17b": "Qwen/Qwen3.5-397B-A17B",
"qwen3.5-35b-a3b": "Qwen/Qwen3.5-35B-A3B",
"qwen3.5-35b-a3b-fp8": "Qwen/Qwen3.5-35B-A3B-FP8",
"qwen3.5-35b-fp8": "Qwen/Qwen3.5-35B-A3B-FP8",
"mixtral-8x7b": "mistralai/Mixtral-8x7B-Instruct-v0.1",
"mistral-7b": "mistralai/Mistral-7B-v0.1",
"mixtral-8x22b": "mistralai/Mixtral-8x22B-Instruct-v0.1",
Expand Down
2 changes: 2 additions & 0 deletions tests/unit/configs_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,6 +246,8 @@ def test_mistral_configs(config_file):
os.path.join(CONFIGS_DIR, "models", "qwen3-480b-a35b.yml"),
os.path.join(CONFIGS_DIR, "models", "qwen3-next-80b-a3b.yml"),
os.path.join(CONFIGS_DIR, "models", "qwen3-omni-30b-a3b.yml"),
os.path.join(CONFIGS_DIR, "models", "qwen3.5-35b-a3b.yml"),
os.path.join(CONFIGS_DIR, "models", "qwen3.5-35b-a3b-fp8.yml"),
]


Expand Down
Loading