Models / qwen3_5_moe

Qwen3.6-35B-A3B-NVFP4: GPU memory to fine-tune

Fine-tuning Qwen3.6-35B-A3B-NVFP4 at batch 1 and 2048 tokens takes about 688 GB of GPU memory for a full fine-tune (at least 6 x H200 sharded), and 81.5 GB with LoRA r=16 on a bf16 base (fits one H200 141 GB). Estimated from the model's structure, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 10.3M downloads on Hugging Face.

Our count against the checkpoint

The left number comes from the graph. The right one is the number of scalars in the published weight files. Nothing on this page was tuned to make them agree.

Derived from structure

35.29B

35,287,636,864 parameters

In the published checkpoint

18.68B

18,683,860,336 scalars · safetensors.total, read 2026-09-06

Delta

+88.9%

quantized This checkpoint is stored quantized (modelopt). The tensor count in the file counts stored elements under a packing scheme, not logical parameters, so the two numbers below are not measuring the same thing in either direction.

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32, 2048 tokens over 50,000 samples (assumed; no dataset attached).

Layers

260

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$242.86

Memory: what fits on which card

Largest batch per card. Training is sized at 2048 tokens per sample, serving at 4096 tokens per sequence.

Full fine-tune, batch 1 at 2048 tokens

688 GB

LoRA r=16, batch 1 at 2048 tokens

81.5 GB

Serve: bf16 weights

70.6 GB

Serve: KV cache, 4096 tokens

84 MB

Serve: recurrent state

33 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
L4 24 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
A100 40 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
A100 / H100 80 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
H200 141 GB	does not fit at batch 1	batch 4	batch 445
Sharded, by memory alone	at least 6 x H200 or 10 x 80 GB

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 262144), AdamW with bf16 autocast, SDPA attention (no score matrix stored), 1.15x headroom on every card. Full fine-tune holds weights, gradients and optimizer state (1.2 x 16 bytes per parameter) plus activations. LoRA is r=16 on every linear layer but the head, base frozen in bf16 (2 bytes per parameter), adapters at full AdamW state, same activations. A two-card split is a contiguous layer cut (pipeline parallel), computed only where one card cannot hold batch 1. "Sharded" is a floor: the batch-1 peak with headroom divided across cards as if it split perfectly evenly (FSDP / ZeRO-3 for training, tensor parallel for serving), with no communication buffers counted. Serving: bf16 weights plus the KV cache of 4096 tokens per sequence plus the fixed recurrent state of the SSM / linear-attention layers; prefill working set, CUDA context and framework reserve not counted.

These are estimates from the graph, not measurements. Against measured runs, held out by architecture, the estimate flagged 113 of the 131 jobs that did not fit their card and said "fits" on the other 18. Every measured run.

Structure

264 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 262144 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 262144
2	Embedding	Embedding	1 × 262144 × 2048
3	RoPE	RoPE	1 × 262144 × 2048
4	Vision input	Input	3 × 224 × 224
5	PatchEmbed	Patch Embed	196 × 1152
6	Patch_Position_Embedding	Learned Pos Embed	196 × 1152
7	Vision encoder (internals not in config)	Projection	196 × 2048
8	Vision tokens	Reshape	1 × 196 × 2048
9	Multimodal fusion (concat tokens)	Concatenate	1 × 262340 × 2048
10	RMSNorm_1_1	RMSNorm	1 × 262340 × 2048
11	DeltaNet_1	Gated DeltaNet	1 × 262340 × 2048
12	Add_1_attn	Add	1 × 262340 × 2048
13	RMSNorm_1_2	RMSNorm	1 × 262340 × 2048
14	MoE_1	MoE Layer	1 × 262340 × 2048
15	Add_1_ffn	Add	1 × 262340 × 2048
16	RMSNorm_2_1	RMSNorm	1 × 262340 × 2048
17	DeltaNet_2	Gated DeltaNet	1 × 262340 × 2048
18	Add_2_attn	Add	1 × 262340 × 2048
19	RMSNorm_2_2	RMSNorm	1 × 262340 × 2048
20	MoE_2	MoE Layer	1 × 262340 × 2048
21	Add_2_ffn	Add	1 × 262340 × 2048
22	RMSNorm_3_1	RMSNorm	1 × 262340 × 2048
23	DeltaNet_3	Gated DeltaNet	1 × 262340 × 2048
24	Add_3_attn	Add	1 × 262340 × 2048
25	RMSNorm_3_2	RMSNorm	1 × 262340 × 2048
26	MoE_3	MoE Layer	1 × 262340 × 2048
27	Add_3_ffn	Add	1 × 262340 × 2048
28	RMSNorm_4_1	RMSNorm	1 × 262340 × 2048
29	Attention_4	Grouped Query Attn	1 × 262340 × 2048
30	Add_4_attn	Add	1 × 262340 × 2048
31	RMSNorm_4_2	RMSNorm	1 × 262340 × 2048
32	MoE_4	MoE Layer	1 × 262340 × 2048
33	Add_4_ffn	Add	1 × 262340 × 2048
34	RMSNorm_5_1	RMSNorm	1 × 262340 × 2048
35	DeltaNet_5	Gated DeltaNet	1 × 262340 × 2048
36	Add_5_attn	Add	1 × 262340 × 2048
37	RMSNorm_5_2	RMSNorm	1 × 262340 × 2048
38	MoE_5	MoE Layer	1 × 262340 × 2048
39	Add_5_ffn	Add	1 × 262340 × 2048
40	RMSNorm_6_1	RMSNorm	1 × 262340 × 2048
41	DeltaNet_6	Gated DeltaNet	1 × 262340 × 2048
42	Add_6_attn	Add	1 × 262340 × 2048
43	RMSNorm_6_2	RMSNorm	1 × 262340 × 2048
44	MoE_6	MoE Layer	1 × 262340 × 2048
45	Add_6_ffn	Add	1 × 262340 × 2048
46	RMSNorm_7_1	RMSNorm	1 × 262340 × 2048
47	DeltaNet_7	Gated DeltaNet	1 × 262340 × 2048
48	Add_7_attn	Add	1 × 262340 × 2048
49	RMSNorm_7_2	RMSNorm	1 × 262340 × 2048
50	MoE_7	MoE Layer	1 × 262340 × 2048
51	Add_7_ffn	Add	1 × 262340 × 2048
52	RMSNorm_8_1	RMSNorm	1 × 262340 × 2048
53	Attention_8	Grouped Query Attn	1 × 262340 × 2048
54	Add_8_attn	Add	1 × 262340 × 2048
55	RMSNorm_8_2	RMSNorm	1 × 262340 × 2048
56	MoE_8	MoE Layer	1 × 262340 × 2048
57	Add_8_ffn	Add	1 × 262340 × 2048
58	RMSNorm_9_1	RMSNorm	1 × 262340 × 2048
59	DeltaNet_9	Gated DeltaNet	1 × 262340 × 2048
60	Add_9_attn	Add	1 × 262340 × 2048
61	RMSNorm_9_2	RMSNorm	1 × 262340 × 2048
62	MoE_9	MoE Layer	1 × 262340 × 2048
63	Add_9_ffn	Add	1 × 262340 × 2048
64	RMSNorm_10_1	RMSNorm	1 × 262340 × 2048
65	DeltaNet_10	Gated DeltaNet	1 × 262340 × 2048
66	Add_10_attn	Add	1 × 262340 × 2048
67	RMSNorm_10_2	RMSNorm	1 × 262340 × 2048
68	MoE_10	MoE Layer	1 × 262340 × 2048
69	Add_10_ffn	Add	1 × 262340 × 2048
70	RMSNorm_11_1	RMSNorm	1 × 262340 × 2048
71	DeltaNet_11	Gated DeltaNet	1 × 262340 × 2048
72	Add_11_attn	Add	1 × 262340 × 2048
73	RMSNorm_11_2	RMSNorm	1 × 262340 × 2048
74	MoE_11	MoE Layer	1 × 262340 × 2048
75	Add_11_ffn	Add	1 × 262340 × 2048
76	RMSNorm_12_1	RMSNorm	1 × 262340 × 2048
77	Attention_12	Grouped Query Attn	1 × 262340 × 2048
78	Add_12_attn	Add	1 × 262340 × 2048
79	RMSNorm_12_2	RMSNorm	1 × 262340 × 2048
80	MoE_12	MoE Layer	1 × 262340 × 2048
81	Add_12_ffn	Add	1 × 262340 × 2048
82	RMSNorm_13_1	RMSNorm	1 × 262340 × 2048
83	DeltaNet_13	Gated DeltaNet	1 × 262340 × 2048
84	Add_13_attn	Add	1 × 262340 × 2048
85	RMSNorm_13_2	RMSNorm	1 × 262340 × 2048
86	MoE_13	MoE Layer	1 × 262340 × 2048
87	Add_13_ffn	Add	1 × 262340 × 2048
88	RMSNorm_14_1	RMSNorm	1 × 262340 × 2048
89	DeltaNet_14	Gated DeltaNet	1 × 262340 × 2048
90	Add_14_attn	Add	1 × 262340 × 2048
91	RMSNorm_14_2	RMSNorm	1 × 262340 × 2048
92	MoE_14	MoE Layer	1 × 262340 × 2048
93	Add_14_ffn	Add	1 × 262340 × 2048
94	RMSNorm_15_1	RMSNorm	1 × 262340 × 2048
95	DeltaNet_15	Gated DeltaNet	1 × 262340 × 2048
96	Add_15_attn	Add	1 × 262340 × 2048
97	RMSNorm_15_2	RMSNorm	1 × 262340 × 2048
98	MoE_15	MoE Layer	1 × 262340 × 2048
99	Add_15_ffn	Add	1 × 262340 × 2048
100	RMSNorm_16_1	RMSNorm	1 × 262340 × 2048
101	Attention_16	Grouped Query Attn	1 × 262340 × 2048
102	Add_16_attn	Add	1 × 262340 × 2048
103	RMSNorm_16_2	RMSNorm	1 × 262340 × 2048
104	MoE_16	MoE Layer	1 × 262340 × 2048
105	Add_16_ffn	Add	1 × 262340 × 2048
106	RMSNorm_17_1	RMSNorm	1 × 262340 × 2048
107	DeltaNet_17	Gated DeltaNet	1 × 262340 × 2048
108	Add_17_attn	Add	1 × 262340 × 2048
109	RMSNorm_17_2	RMSNorm	1 × 262340 × 2048
110	MoE_17	MoE Layer	1 × 262340 × 2048
111	Add_17_ffn	Add	1 × 262340 × 2048
112	RMSNorm_18_1	RMSNorm	1 × 262340 × 2048
113	DeltaNet_18	Gated DeltaNet	1 × 262340 × 2048
114	Add_18_attn	Add	1 × 262340 × 2048
115	RMSNorm_18_2	RMSNorm	1 × 262340 × 2048
116	MoE_18	MoE Layer	1 × 262340 × 2048
117	Add_18_ffn	Add	1 × 262340 × 2048
118	RMSNorm_19_1	RMSNorm	1 × 262340 × 2048
119	DeltaNet_19	Gated DeltaNet	1 × 262340 × 2048
120	Add_19_attn	Add	1 × 262340 × 2048
121	RMSNorm_19_2	RMSNorm	1 × 262340 × 2048
122	MoE_19	MoE Layer	1 × 262340 × 2048
123	Add_19_ffn	Add	1 × 262340 × 2048
124	RMSNorm_20_1	RMSNorm	1 × 262340 × 2048
125	Attention_20	Grouped Query Attn	1 × 262340 × 2048
126	Add_20_attn	Add	1 × 262340 × 2048
127	RMSNorm_20_2	RMSNorm	1 × 262340 × 2048
128	MoE_20	MoE Layer	1 × 262340 × 2048
129	Add_20_ffn	Add	1 × 262340 × 2048
130	RMSNorm_21_1	RMSNorm	1 × 262340 × 2048
131	DeltaNet_21	Gated DeltaNet	1 × 262340 × 2048
132	Add_21_attn	Add	1 × 262340 × 2048
133	RMSNorm_21_2	RMSNorm	1 × 262340 × 2048
134	MoE_21	MoE Layer	1 × 262340 × 2048
135	Add_21_ffn	Add	1 × 262340 × 2048
136	RMSNorm_22_1	RMSNorm	1 × 262340 × 2048
137	DeltaNet_22	Gated DeltaNet	1 × 262340 × 2048
138	Add_22_attn	Add	1 × 262340 × 2048
139	RMSNorm_22_2	RMSNorm	1 × 262340 × 2048
140	MoE_22	MoE Layer	1 × 262340 × 2048
141	Add_22_ffn	Add	1 × 262340 × 2048
142	RMSNorm_23_1	RMSNorm	1 × 262340 × 2048
143	DeltaNet_23	Gated DeltaNet	1 × 262340 × 2048
144	Add_23_attn	Add	1 × 262340 × 2048
145	RMSNorm_23_2	RMSNorm	1 × 262340 × 2048
146	MoE_23	MoE Layer	1 × 262340 × 2048
147	Add_23_ffn	Add	1 × 262340 × 2048
148	RMSNorm_24_1	RMSNorm	1 × 262340 × 2048
149	Attention_24	Grouped Query Attn	1 × 262340 × 2048
150	Add_24_attn	Add	1 × 262340 × 2048
151	RMSNorm_24_2	RMSNorm	1 × 262340 × 2048
152	MoE_24	MoE Layer	1 × 262340 × 2048
153	Add_24_ffn	Add	1 × 262340 × 2048
154	RMSNorm_25_1	RMSNorm	1 × 262340 × 2048
155	DeltaNet_25	Gated DeltaNet	1 × 262340 × 2048
156	Add_25_attn	Add	1 × 262340 × 2048
157	RMSNorm_25_2	RMSNorm	1 × 262340 × 2048
158	MoE_25	MoE Layer	1 × 262340 × 2048
159	Add_25_ffn	Add	1 × 262340 × 2048
160	RMSNorm_26_1	RMSNorm	1 × 262340 × 2048
161	DeltaNet_26	Gated DeltaNet	1 × 262340 × 2048
162	Add_26_attn	Add	1 × 262340 × 2048
163	RMSNorm_26_2	RMSNorm	1 × 262340 × 2048
164	MoE_26	MoE Layer	1 × 262340 × 2048
165	Add_26_ffn	Add	1 × 262340 × 2048
166	RMSNorm_27_1	RMSNorm	1 × 262340 × 2048
167	DeltaNet_27	Gated DeltaNet	1 × 262340 × 2048
168	Add_27_attn	Add	1 × 262340 × 2048
169	RMSNorm_27_2	RMSNorm	1 × 262340 × 2048
170	MoE_27	MoE Layer	1 × 262340 × 2048
171	Add_27_ffn	Add	1 × 262340 × 2048
172	RMSNorm_28_1	RMSNorm	1 × 262340 × 2048
173	Attention_28	Grouped Query Attn	1 × 262340 × 2048
174	Add_28_attn	Add	1 × 262340 × 2048
175	RMSNorm_28_2	RMSNorm	1 × 262340 × 2048
176	MoE_28	MoE Layer	1 × 262340 × 2048
177	Add_28_ffn	Add	1 × 262340 × 2048
178	RMSNorm_29_1	RMSNorm	1 × 262340 × 2048
179	DeltaNet_29	Gated DeltaNet	1 × 262340 × 2048
180	Add_29_attn	Add	1 × 262340 × 2048
181	RMSNorm_29_2	RMSNorm	1 × 262340 × 2048
182	MoE_29	MoE Layer	1 × 262340 × 2048
183	Add_29_ffn	Add	1 × 262340 × 2048
184	RMSNorm_30_1	RMSNorm	1 × 262340 × 2048
185	DeltaNet_30	Gated DeltaNet	1 × 262340 × 2048
186	Add_30_attn	Add	1 × 262340 × 2048
187	RMSNorm_30_2	RMSNorm	1 × 262340 × 2048
188	MoE_30	MoE Layer	1 × 262340 × 2048
189	Add_30_ffn	Add	1 × 262340 × 2048
190	RMSNorm_31_1	RMSNorm	1 × 262340 × 2048
191	DeltaNet_31	Gated DeltaNet	1 × 262340 × 2048
192	Add_31_attn	Add	1 × 262340 × 2048
193	RMSNorm_31_2	RMSNorm	1 × 262340 × 2048
194	MoE_31	MoE Layer	1 × 262340 × 2048
195	Add_31_ffn	Add	1 × 262340 × 2048
196	RMSNorm_32_1	RMSNorm	1 × 262340 × 2048
197	Attention_32	Grouped Query Attn	1 × 262340 × 2048
198	Add_32_attn	Add	1 × 262340 × 2048
199	RMSNorm_32_2	RMSNorm	1 × 262340 × 2048
200	MoE_32	MoE Layer	1 × 262340 × 2048
201	Add_32_ffn	Add	1 × 262340 × 2048
202	RMSNorm_33_1	RMSNorm	1 × 262340 × 2048
203	DeltaNet_33	Gated DeltaNet	1 × 262340 × 2048
204	Add_33_attn	Add	1 × 262340 × 2048
205	RMSNorm_33_2	RMSNorm	1 × 262340 × 2048
206	MoE_33	MoE Layer	1 × 262340 × 2048
207	Add_33_ffn	Add	1 × 262340 × 2048
208	RMSNorm_34_1	RMSNorm	1 × 262340 × 2048
209	DeltaNet_34	Gated DeltaNet	1 × 262340 × 2048
210	Add_34_attn	Add	1 × 262340 × 2048
211	RMSNorm_34_2	RMSNorm	1 × 262340 × 2048
212	MoE_34	MoE Layer	1 × 262340 × 2048
213	Add_34_ffn	Add	1 × 262340 × 2048
214	RMSNorm_35_1	RMSNorm	1 × 262340 × 2048
215	DeltaNet_35	Gated DeltaNet	1 × 262340 × 2048
216	Add_35_attn	Add	1 × 262340 × 2048
217	RMSNorm_35_2	RMSNorm	1 × 262340 × 2048
218	MoE_35	MoE Layer	1 × 262340 × 2048
219	Add_35_ffn	Add	1 × 262340 × 2048
220	RMSNorm_36_1	RMSNorm	1 × 262340 × 2048
221	Attention_36	Grouped Query Attn	1 × 262340 × 2048
222	Add_36_attn	Add	1 × 262340 × 2048
223	RMSNorm_36_2	RMSNorm	1 × 262340 × 2048
224	MoE_36	MoE Layer	1 × 262340 × 2048
225	Add_36_ffn	Add	1 × 262340 × 2048
226	RMSNorm_37_1	RMSNorm	1 × 262340 × 2048
227	DeltaNet_37	Gated DeltaNet	1 × 262340 × 2048
228	Add_37_attn	Add	1 × 262340 × 2048
229	RMSNorm_37_2	RMSNorm	1 × 262340 × 2048
230	MoE_37	MoE Layer	1 × 262340 × 2048
231	Add_37_ffn	Add	1 × 262340 × 2048
232	RMSNorm_38_1	RMSNorm	1 × 262340 × 2048
233	DeltaNet_38	Gated DeltaNet	1 × 262340 × 2048
234	Add_38_attn	Add	1 × 262340 × 2048
235	RMSNorm_38_2	RMSNorm	1 × 262340 × 2048
236	MoE_38	MoE Layer	1 × 262340 × 2048
237	Add_38_ffn	Add	1 × 262340 × 2048
238	RMSNorm_39_1	RMSNorm	1 × 262340 × 2048
239	DeltaNet_39	Gated DeltaNet	1 × 262340 × 2048
240	Add_39_attn	Add	1 × 262340 × 2048
241	RMSNorm_39_2	RMSNorm	1 × 262340 × 2048
242	MoE_39	MoE Layer	1 × 262340 × 2048
243	Add_39_ffn	Add	1 × 262340 × 2048
244	RMSNorm_40_1	RMSNorm	1 × 262340 × 2048
245	Attention_40	Grouped Query Attn	1 × 262340 × 2048
246	Add_40_attn	Add	1 × 262340 × 2048
247	RMSNorm_40_2	RMSNorm	1 × 262340 × 2048
248	MoE_40	MoE Layer	1 × 262340 × 2048
249	Add_40_ffn	Add	1 × 262340 × 2048
250	MTP_1_HiddenNorm	RMSNorm	1 × 262340 × 2048
251	MTP_1_EmbeddingNorm	RMSNorm	1 × 262340 × 2048
252	MTP_1_Concat	Concatenate	1 × 262340 × 4096
253	MTP_1_Projection	Linear	1 × 262340 × 2048
254	RMSNorm_MTP_1_1	RMSNorm	1 × 262340 × 2048
255	Attention_MTP_1	Grouped Query Attn	1 × 262340 × 2048
256	Add_MTP_1_attn	Add	1 × 262340 × 2048
257	RMSNorm_MTP_1_2	RMSNorm	1 × 262340 × 2048
258	MoE_MTP_1	MoE Layer	1 × 262340 × 2048
259	Add_MTP_1_ffn	Add	1 × 262340 × 2048
260	MTP_1_Norm	RMSNorm	1 × 262340 × 2048
261	Final_RMSNorm	RMSNorm	1 × 262340 × 2048
262	LM_Head	Linear	1 × 262340 × 248320
263	Output	Output	1 × 262340 × 248320
264	MTP_1_Output	Output	1 × 262340 × 2048

What the verifier says

warn"RoPE" receives input but its output is not connected. This layer will be unreachable in the forward pass. Fix: Connect the output forward, or add an Output node if this is the final layer.
dead-end

infoMoE layers require an auxiliary router z-loss + load-balance loss during training to prevent expert collapse. This is not visible in the architecture diagram but must be in the training loop. Applies to all 41: MoE_1, MoE_2, MoE_3, MoE_4, MoE_5, MoE_6, MoE_7, MoE_8, MoE_9, MoE_10, MoE_11, MoE_12, MoE_13, MoE_14, MoE_15, MoE_16, MoE_17, MoE_18, MoE_19, MoE_20, MoE_21, MoE_22, MoE_23, MoE_24, MoE_25, MoE_26, MoE_27, MoE_28, MoE_29, MoE_30, MoE_31, MoE_32, MoE_33, MoE_34, MoE_35, MoE_36, MoE_37, MoE_38, MoE_39, MoE_40, MoE_MTP_1. Fix: Add a note on these layers. Typical aux_loss coefficient: 1e-2 (Mixtral/Switch Transformer).
moe-no-aux-loss

infoAt 11 stacked attention layers, residual-branch outputs add up; unscaled init lets activation variance grow with depth. GPT-2/LLaMA-family models scale the residual projections by depth (N(0, 0.02 / √(2L))). Fix: Scale residual output projections by depth: nn.init.normal_(w, std=0.02 / math.sqrt(2 * n_layers))
deep-attention-default-init

What changed upstream

The change tracker follows nvidia/Qwen3.6-35B-A3B-NVFP4: every revision it has read, what changed in the config, tokenizer, chat template or weights between them, and an OSV advisory per change.

Revision history and changes Badge JSON OSV advisories All tracked changes

Do this to your own model

Same numbers, on a model in your repo, in one command. No account.

pip install neurarch-trace
neurarch-trace nvidia/Qwen3.6-35B-A3B-NVFP4 --plan --share

Paste a model instead Markdown twin Index JSON POST /api/v1/plan