Add matched CR10 strict M1 inference checkpoints at 3k, 5k and 10k

.gitattributes

@@ -117,3 +117,9 @@ c1-1/cr10_staged_cosine/packed_temporal_grid/iter_000004000/optim_compressor/.me
 c1-1/cr10_staged_cosine/packed_temporal_grid/iter_000004000/optim_compressor/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
 c3-3/cr10_fullprefix_gradctxfix/iter_000002000/model/.metadata filter=lfs diff=lfs merge=lfs -text
 c3-3/cr10_fullprefix_gradctxfix/iter_000002000/model/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
+c3-3/cr10_fullprefix_gradctxfix/iter_000003000/model/.metadata filter=lfs diff=lfs merge=lfs -text
+c3-3/cr10_fullprefix_gradctxfix/iter_000003000/model/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
+c3-3/cr10_fullprefix_gradctxfix/iter_000005000/model/.metadata filter=lfs diff=lfs merge=lfs -text
+c3-3/cr10_fullprefix_gradctxfix/iter_000005000/model/__0_0.distcp filter=lfs diff=lfs merge=lfs -text
+c3-3/cr10_fullprefix_gradctxfix/iter_000010000/model/.metadata filter=lfs diff=lfs merge=lfs -text
+c3-3/cr10_fullprefix_gradctxfix/iter_000010000/model/__0_0.distcp filter=lfs diff=lfs merge=lfs -text

c3-3/cr10_fullprefix_gradctxfix/iter_000003000/README.md

@@ -0,0 +1,69 @@
+# CR10 strict M=1: gradient-fixed c3-3 compressor, iteration 3000
+
+This is iteration 3000 of the same gradient-fixed CR10 strict M=1 training run as the shared 2k checkpoint. This upload makes no new performance claim.
+It is not the older c3-3 CR4/packed checkpoint or the c1-1 CR10 checkpoint.
+The run name ends in `raw_10k`, but the supplied checkpoint is iteration **3000**.
+
+## Code
+
+Repository: https://github.com/a-little-hoof/video-compress
+
+Pinned commit: `757194f78bad10f50d660d14342a5f18b82470f0`
+
+Branch: `exp/relax-official-shared-c3-20260909`
+
+The GitHub repository requires its own access authorization. Hugging Face read
+access does not automatically grant GitHub access.
+
+Entry point: `scripts/context_compression/generate_mbencht_c3_3_continuations.py`.
+The M=1 implementation is `generate_strict_relax_history_slot`; its exact method
+identifier is `c3_relax_slot_h1_cr10_matched_interval_replacement`.
+Raw matched control: `c3_relax_slot_h1_raw_control`.
+Released-policy adaptation: `c3_relax_forcing4_official_shared_frame`.
+The runtime is `rcm/inference/wan2pt1_t2v_causal_extrapolation_infer.py`.
+
+## Required separate generator
+
+Repository: `worstcoder/rcm-Wan`
+
+Revision: `142d77a331f1516fea772607c1bf25a2785a2505`
+
+Filename: `Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step4.pt`
+
+SHA256: `cc53e27e4501f58f43d1ca01dc2e5dde4df946ce5c85ab79695a902d929df008`
+
+Do not substitute the similarly named TF-sCM-init generator. The VAE and text
+encoder are `Wan2.1_VAE.pth` and `models_t5_umt5-xxl-enc-bf16.pth` in
+`worstcoder/Wan`, revision `18554de6b48df8ab2ae274972a076b27ad5cb602`.
+
+## Loading
+
+Download this entire checkpoint directory, including hidden `model/.metadata`
+and all `model/*.distcp` files. Pass its parent-of-model directory to
+`--packed_compressor_ckpt`, with `--checkpoint_iteration 3000 --memory_ratio 0.1`.
+Use `--model_size 1.3B --max_memory_tokens_per_frame 1024`.
+The pinned `generate_compressed_context.py::load_compressor_weights` supports
+this distributed-checkpoint layout. Optimizer state is not required for inference
+and is not included. This package is not a resumable training checkpoint.
+
+## What M=1 means
+
+At 832x480, active past context is two raw sink frames, one selected compressed
+1560-token frame-equivalent unit, and one raw recent frame: 6240 past KV tokens
+per layer. The current c3 chunk adds three latent frames. CR10 is the compressor
+ratio, not the number of active memory slots.
+
+Compress each complete 21-latent-frame block once. Contextualize the full
+compressed archive bidirectionally before selecting the one active unit;
+retain residual tokens for contextualization but not as selectable full units.
+The historical M=1 method selects independently per layer and freezes selected
+history KV between 21-frame boundaries. Raw bootstrap covers [0,42); compressed memory first becomes active at latent frame 42.
+
+The released Relax adaptation differs in layer-shared selection, candidate
+grid, prototype reduction, reservoir policy, and three-frame refresh cadence.
+Changing those is an inference ablation, not an identical reproduction of the
+historical M=1 result. Keep the generator, checkpoint, prompts/noise, active
+context budget, and scoring protocol matched when testing individual changes.
+
+File hashes are supplied in `sha256_manifest.json`. No credentials, private
+launchers, or new training/evaluation results are included in this package.

c3-3/cr10_fullprefix_gradctxfix/iter_000003000/inference_config.json

@@ -0,0 +1,18 @@
+{
+  "checkpoint_iteration": 3000,
+  "checkpoint_layout": "pytorch_distributed_checkpoint",
+  "fps": 16,
+  "generation_protocol": "official_continuous_seed0_v1",
+  "max_memory_tokens_per_frame": 1024,
+  "memory_ratio": 0.1,
+  "method": "c3_relax_slot_h1_cr10_matched_interval_replacement",
+  "mid_t": [
+    "15/16",
+    "5/6",
+    "5/8"
+  ],
+  "model_size": "1.3B",
+  "num_steps": 4,
+  "seed": 0,
+  "sigma_max": 1600
+}

c3-3/cr10_fullprefix_gradctxfix/iter_000003000/model/.metadata

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ec0f6598993b4d8007bf860433eae6835879411fdb0f990f403f49f0a952811
+size 481484

c3-3/cr10_fullprefix_gradctxfix/iter_000003000/model/__0_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14716eccb23e91ad430d73c812aa49c4027eadd4d4ea280a06dc5171fc23ea7e
+size 5681533593

c3-3/cr10_fullprefix_gradctxfix/iter_000003000/sha256_manifest.json

@@ -0,0 +1,35 @@
+{
+  "files": [
+    {
+      "path": "model/.metadata",
+      "sha256": "0ec0f6598993b4d8007bf860433eae6835879411fdb0f990f403f49f0a952811",
+      "size": 481484
+    },
+    {
+      "path": "model/__0_0.distcp",
+      "sha256": "14716eccb23e91ad430d73c812aa49c4027eadd4d4ea280a06dc5171fc23ea7e",
+      "size": 5681533593
+    },
+    {
+      "path": "model/__1_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__2_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__3_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    }
+  ],
+  "inference_only": true,
+  "iteration": 3000,
+  "schema": "c3-strict-m1-checkpoint-handoff/v1",
+  "source_commit": "757194f78bad10f50d660d14342a5f18b82470f0",
+  "training_run": "fullprefix_memoryonly_cr10_c3_3_gradctxfix_gbs8_lr1e4_stagedcosine_clip5_raw_10k",
+  "variant": "c3-3 CR10 full-prefix memory-only gradient-fixed"
+}

c3-3/cr10_fullprefix_gradctxfix/iter_000005000/README.md

@@ -0,0 +1,69 @@
+# CR10 strict M=1: gradient-fixed c3-3 compressor, iteration 5000
+
+This is iteration 5000 of the same gradient-fixed CR10 strict M=1 training run as the shared 2k checkpoint. This upload makes no new performance claim.
+It is not the older c3-3 CR4/packed checkpoint or the c1-1 CR10 checkpoint.
+The run name ends in `raw_10k`, but the supplied checkpoint is iteration **5000**.
+
+## Code
+
+Repository: https://github.com/a-little-hoof/video-compress
+
+Pinned commit: `757194f78bad10f50d660d14342a5f18b82470f0`
+
+Branch: `exp/relax-official-shared-c3-20260909`
+
+The GitHub repository requires its own access authorization. Hugging Face read
+access does not automatically grant GitHub access.
+
+Entry point: `scripts/context_compression/generate_mbencht_c3_3_continuations.py`.
+The M=1 implementation is `generate_strict_relax_history_slot`; its exact method
+identifier is `c3_relax_slot_h1_cr10_matched_interval_replacement`.
+Raw matched control: `c3_relax_slot_h1_raw_control`.
+Released-policy adaptation: `c3_relax_forcing4_official_shared_frame`.
+The runtime is `rcm/inference/wan2pt1_t2v_causal_extrapolation_infer.py`.
+
+## Required separate generator
+
+Repository: `worstcoder/rcm-Wan`
+
+Revision: `142d77a331f1516fea772607c1bf25a2785a2505`
+
+Filename: `Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step4.pt`
+
+SHA256: `cc53e27e4501f58f43d1ca01dc2e5dde4df946ce5c85ab79695a902d929df008`
+
+Do not substitute the similarly named TF-sCM-init generator. The VAE and text
+encoder are `Wan2.1_VAE.pth` and `models_t5_umt5-xxl-enc-bf16.pth` in
+`worstcoder/Wan`, revision `18554de6b48df8ab2ae274972a076b27ad5cb602`.
+
+## Loading
+
+Download this entire checkpoint directory, including hidden `model/.metadata`
+and all `model/*.distcp` files. Pass its parent-of-model directory to
+`--packed_compressor_ckpt`, with `--checkpoint_iteration 5000 --memory_ratio 0.1`.
+Use `--model_size 1.3B --max_memory_tokens_per_frame 1024`.
+The pinned `generate_compressed_context.py::load_compressor_weights` supports
+this distributed-checkpoint layout. Optimizer state is not required for inference
+and is not included. This package is not a resumable training checkpoint.
+
+## What M=1 means
+
+At 832x480, active past context is two raw sink frames, one selected compressed
+1560-token frame-equivalent unit, and one raw recent frame: 6240 past KV tokens
+per layer. The current c3 chunk adds three latent frames. CR10 is the compressor
+ratio, not the number of active memory slots.
+
+Compress each complete 21-latent-frame block once. Contextualize the full
+compressed archive bidirectionally before selecting the one active unit;
+retain residual tokens for contextualization but not as selectable full units.
+The historical M=1 method selects independently per layer and freezes selected
+history KV between 21-frame boundaries. Raw bootstrap covers [0,42); compressed memory first becomes active at latent frame 42.
+
+The released Relax adaptation differs in layer-shared selection, candidate
+grid, prototype reduction, reservoir policy, and three-frame refresh cadence.
+Changing those is an inference ablation, not an identical reproduction of the
+historical M=1 result. Keep the generator, checkpoint, prompts/noise, active
+context budget, and scoring protocol matched when testing individual changes.
+
+File hashes are supplied in `sha256_manifest.json`. No credentials, private
+launchers, or new training/evaluation results are included in this package.

c3-3/cr10_fullprefix_gradctxfix/iter_000005000/inference_config.json

@@ -0,0 +1,18 @@
+{
+  "checkpoint_iteration": 5000,
+  "checkpoint_layout": "pytorch_distributed_checkpoint",
+  "fps": 16,
+  "generation_protocol": "official_continuous_seed0_v1",
+  "max_memory_tokens_per_frame": 1024,
+  "memory_ratio": 0.1,
+  "method": "c3_relax_slot_h1_cr10_matched_interval_replacement",
+  "mid_t": [
+    "15/16",
+    "5/6",
+    "5/8"
+  ],
+  "model_size": "1.3B",
+  "num_steps": 4,
+  "seed": 0,
+  "sigma_max": 1600
+}

c3-3/cr10_fullprefix_gradctxfix/iter_000005000/model/.metadata

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd54f54b707454fc216033835825ae385851a22bd7652f11a05d91ca99eae3c9
+size 481484

c3-3/cr10_fullprefix_gradctxfix/iter_000005000/model/__0_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9d1112a90c465f384069c14b6b6c152993d826bc86c863ce3263ca3a121a420
+size 5681533593

c3-3/cr10_fullprefix_gradctxfix/iter_000005000/sha256_manifest.json

@@ -0,0 +1,35 @@
+{
+  "files": [
+    {
+      "path": "model/.metadata",
+      "sha256": "cd54f54b707454fc216033835825ae385851a22bd7652f11a05d91ca99eae3c9",
+      "size": 481484
+    },
+    {
+      "path": "model/__0_0.distcp",
+      "sha256": "a9d1112a90c465f384069c14b6b6c152993d826bc86c863ce3263ca3a121a420",
+      "size": 5681533593
+    },
+    {
+      "path": "model/__1_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__2_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__3_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    }
+  ],
+  "inference_only": true,
+  "iteration": 5000,
+  "schema": "c3-strict-m1-checkpoint-handoff/v1",
+  "source_commit": "757194f78bad10f50d660d14342a5f18b82470f0",
+  "training_run": "fullprefix_memoryonly_cr10_c3_3_gradctxfix_gbs8_lr1e4_stagedcosine_clip5_raw_10k",
+  "variant": "c3-3 CR10 full-prefix memory-only gradient-fixed"
+}

c3-3/cr10_fullprefix_gradctxfix/iter_000010000/README.md

@@ -0,0 +1,69 @@
+# CR10 strict M=1: gradient-fixed c3-3 compressor, iteration 10000
+
+This is iteration 10000 of the same gradient-fixed CR10 strict M=1 training run as the shared 2k checkpoint. This upload makes no new performance claim.
+It is not the older c3-3 CR4/packed checkpoint or the c1-1 CR10 checkpoint.
+The run name ends in `raw_10k`, but the supplied checkpoint is iteration **10000**.
+
+## Code
+
+Repository: https://github.com/a-little-hoof/video-compress
+
+Pinned commit: `757194f78bad10f50d660d14342a5f18b82470f0`
+
+Branch: `exp/relax-official-shared-c3-20260909`
+
+The GitHub repository requires its own access authorization. Hugging Face read
+access does not automatically grant GitHub access.
+
+Entry point: `scripts/context_compression/generate_mbencht_c3_3_continuations.py`.
+The M=1 implementation is `generate_strict_relax_history_slot`; its exact method
+identifier is `c3_relax_slot_h1_cr10_matched_interval_replacement`.
+Raw matched control: `c3_relax_slot_h1_raw_control`.
+Released-policy adaptation: `c3_relax_forcing4_official_shared_frame`.
+The runtime is `rcm/inference/wan2pt1_t2v_causal_extrapolation_infer.py`.
+
+## Required separate generator
+
+Repository: `worstcoder/rcm-Wan`
+
+Revision: `142d77a331f1516fea772607c1bf25a2785a2505`
+
+Filename: `Causal_rCM_Wan2.1_T2V_1.3B_480p_TF-dCM-init_SF-DMD_c3-3_step4.pt`
+
+SHA256: `cc53e27e4501f58f43d1ca01dc2e5dde4df946ce5c85ab79695a902d929df008`
+
+Do not substitute the similarly named TF-sCM-init generator. The VAE and text
+encoder are `Wan2.1_VAE.pth` and `models_t5_umt5-xxl-enc-bf16.pth` in
+`worstcoder/Wan`, revision `18554de6b48df8ab2ae274972a076b27ad5cb602`.
+
+## Loading
+
+Download this entire checkpoint directory, including hidden `model/.metadata`
+and all `model/*.distcp` files. Pass its parent-of-model directory to
+`--packed_compressor_ckpt`, with `--checkpoint_iteration 10000 --memory_ratio 0.1`.
+Use `--model_size 1.3B --max_memory_tokens_per_frame 1024`.
+The pinned `generate_compressed_context.py::load_compressor_weights` supports
+this distributed-checkpoint layout. Optimizer state is not required for inference
+and is not included. This package is not a resumable training checkpoint.
+
+## What M=1 means
+
+At 832x480, active past context is two raw sink frames, one selected compressed
+1560-token frame-equivalent unit, and one raw recent frame: 6240 past KV tokens
+per layer. The current c3 chunk adds three latent frames. CR10 is the compressor
+ratio, not the number of active memory slots.
+
+Compress each complete 21-latent-frame block once. Contextualize the full
+compressed archive bidirectionally before selecting the one active unit;
+retain residual tokens for contextualization but not as selectable full units.
+The historical M=1 method selects independently per layer and freezes selected
+history KV between 21-frame boundaries. Raw bootstrap covers [0,42); compressed memory first becomes active at latent frame 42.
+
+The released Relax adaptation differs in layer-shared selection, candidate
+grid, prototype reduction, reservoir policy, and three-frame refresh cadence.
+Changing those is an inference ablation, not an identical reproduction of the
+historical M=1 result. Keep the generator, checkpoint, prompts/noise, active
+context budget, and scoring protocol matched when testing individual changes.
+
+File hashes are supplied in `sha256_manifest.json`. No credentials, private
+launchers, or new training/evaluation results are included in this package.

c3-3/cr10_fullprefix_gradctxfix/iter_000010000/inference_config.json

@@ -0,0 +1,18 @@
+{
+  "checkpoint_iteration": 10000,
+  "checkpoint_layout": "pytorch_distributed_checkpoint",
+  "fps": 16,
+  "generation_protocol": "official_continuous_seed0_v1",
+  "max_memory_tokens_per_frame": 1024,
+  "memory_ratio": 0.1,
+  "method": "c3_relax_slot_h1_cr10_matched_interval_replacement",
+  "mid_t": [
+    "15/16",
+    "5/6",
+    "5/8"
+  ],
+  "model_size": "1.3B",
+  "num_steps": 4,
+  "seed": 0,
+  "sigma_max": 1600
+}

c3-3/cr10_fullprefix_gradctxfix/iter_000010000/model/.metadata

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24b27d546aa581db0b01c60c124a82979530a43ae6ad7e1e514814365bcd3fad
+size 481484

c3-3/cr10_fullprefix_gradctxfix/iter_000010000/model/__0_0.distcp

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7357badbf7aaac6e93475bfe540c1923af3320898e2c1a7967119e7534795987
+size 5681533593

c3-3/cr10_fullprefix_gradctxfix/iter_000010000/sha256_manifest.json

@@ -0,0 +1,35 @@
+{
+  "files": [
+    {
+      "path": "model/.metadata",
+      "sha256": "24b27d546aa581db0b01c60c124a82979530a43ae6ad7e1e514814365bcd3fad",
+      "size": 481484
+    },
+    {
+      "path": "model/__0_0.distcp",
+      "sha256": "7357badbf7aaac6e93475bfe540c1923af3320898e2c1a7967119e7534795987",
+      "size": 5681533593
+    },
+    {
+      "path": "model/__1_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__2_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    },
+    {
+      "path": "model/__3_0.distcp",
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "size": 0
+    }
+  ],
+  "inference_only": true,
+  "iteration": 10000,
+  "schema": "c3-strict-m1-checkpoint-handoff/v1",
+  "source_commit": "757194f78bad10f50d660d14342a5f18b82470f0",
+  "training_run": "fullprefix_memoryonly_cr10_c3_3_gradctxfix_gbs8_lr1e4_stagedcosine_clip5_raw_10k",
+  "variant": "c3-3 CR10 full-prefix memory-only gradient-fixed"
+}